doi
stringlengths 9
10
| chunk-id
stringlengths 1
4
| chunk
stringlengths 1
1.57k
| id
stringlengths 9
10
| title
stringlengths 10
127
| summary
stringlengths 581
1.92k
| source
stringlengths 30
31
| authors
sequence | categories
sequence | comment
stringclasses 28
values | journal_ref
stringclasses 1
value | primary_category
stringclasses 6
values | published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1903.00089 | 9 | implemented in Tensorflow-Lingvo (Shen et al.,
2019).
In all cases we report test results for the checkpoint that performed best on the development set
in terms of BLEU. For the multilingual models we
create a development set that includes examples
we uniformly sample from a concatenation of all
the individual language pair development sets, resulting in 13k development examples per model.
Another important detail regarding multilingual
training is the batching scheme. In all of our multilingual models we use heterogeneous batching,
where each batch contains examples which are
uniformly sampled from a concatenation of all the
language pairs the model is trained on. Specifically, we use batches of 64 examples for sequences
shorter than 69 tokens and batches of 16 examples for longer sequences. We did not use oversampling as the dataset is relatively small.
2.3 Results
We use tokenized BLEU in order to be comparable with Neubig and Hu (2018). Table 1 showsAz-En Be-En Gl-En Sk-En Avg.
# of examples 5.9k 4.5k 10k 61k 20.3k
Neubig & Hu 18
baselines 2.7 2.8 16.2 24 11.42 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 10 | Neubig & Hu 18
baselines 2.7 2.8 16.2 24 11.42
many-to-one 11.7 18.3 29.1 28.3 21.85
Wang et al. 18 11.82 18.71 30.3 28.77 22.4
Ours
many-to-one 11.24 18.28 28.63 26.78 21.23
many-to-many 12.78 21.73 30.65 29.54 23.67
Table 1: X !En test BLEU on the TED Talks corpus,
for the language pairs from Neubig and Hu (2018)
Ar-En De-En He-En It-En Avg.
# of examples 213k 167k 211k 203k 198.5k
baselines 27.84 30.5 34.37 33.64 31.59
many-to-one 25.93 28.87 30.19 32.42 29.35
many-to-many 28.32 32.97 33.18 35.14 32.4
Table 2: X !En test BLEU on the TED Talks corpus,
for language pairs with more than 167k examples
updates ar_many2one ro_many2one nl_many2one de_many2one it_many2one ar de it nl
0 0 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 11 | updates ar_many2one ro_many2one nl_many2one de_many2one it_many2one ar de it nl
0 0
49148 0.2839321792
51751 0.2832087576
53492 0.2823074162
56111 0.2822623849
57843 0.2835516334
59595 0.2890901268
61344 0.2845613658
63938 0.2924402654
65676 0.2828268111
67401 0.2909407914
70019 0.2906394601
71770 0.283375144
73479 0.2862167656
75201 0.2850508094
77780 0.2772497535
79488 0.286601305
81201 0.2964281738
82895 0.2811564505
84594 0.2892799973
86301 0.2811136246
88849 0.2873187661
90551 0.287561506
93101 0.2806210518
95667 0.2852652967
97386 0.286067158
99069 0.285282582
101647 0.2873030901 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 12 | 95667 0.2852652967
97386 0.286067158
99069 0.285282582
101647 0.2873030901
103358 0.2812287509
105932 0.2841886282
107656 0.2752591372
109350 0.2852451503
111069 0.2802457809
113631 0.2839408815
115325 0.2698247731
117038 0.2862476707
119599 0.2838998437
121302 0.2805997431
123015 0.2826767564
Figure 1: Development BLEU on
fIt,Ro,Nl,De,Ar g!En vs. training BLEU for the
many-to-one and many-to-many models. Best viewed
in color.
the results of our experiments when evaluating
on the same language pairs as they did. The results under “Neubig & Hu 18” are their bilingual baselines and their best many-to-one models.
Their many-to-one models use similar-languageregularization, i.e. fine-tuning a pre-trained manyto-one model with data from the language pair of
interest together with data from a language pair | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 13 | interest together with data from a language pair
that has a typologically-similar source language
and more training data (i.e. Russian and Belarusian, Turkish and Azerbaijani). The results under
“Ours” are our many-to-one and many-to-many
models we trained identically in terms of model
architecture and hyper-parameters.
We first note that our many-to-many model outperforms all other models when translating into
English, with 1.82 BLEU improvement (when averaged across the four language pairs) over the
best fine-tuned many-to-one models of Neubig
and Hu (2018) and 2.44 BLEU improvement over
our many-to-one model when averaged across the
four low-resource language pairs (Table 1). This
is surprising as it uses the same X !En data,
model architecture and capacity as our many-toone model, while handling a heavier burden since
it also supports 58 additional translation tasks
(from English into58 languages). Our models also
outperform the more complex models of Wang
et al. (2019) which use ”Soft Decoupled Encoding” for the input tokens, while our models use a
simple subword segmentation. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 14 | et al. (2019) which use ”Soft Decoupled Encoding” for the input tokens, while our models use a
simple subword segmentation.
One possible explanation is that the many-toone model overfits the English side of the corpus
as it is multi-way-parallel: in such setting the English sentences are overlapping across the different language pairs, making it much easier for the
model to memorize the training set instead of generalizing (when enough capacity is available). On
the other hand, the many-to-many model is trained
on additional target languages other than English,
which can act as regularizers for the X !En tasks,
reducing such overfitting.
To further illustrate this, Figure 1 tracks the
BLEU scores on the individual development sets
during training for Italian (It), Romanian (Ro),
Dutch (Nl), German (De) and Arabic (Ar) into English (left), together with BLEU scores on a subset of the training set for each model. We can
see that while the many-to-one model degrades in
performance on the development set, the manyto-many model still improves. Note the large
gap in the many-to-one model between the training set BLEU and the development set BLEU,
which points on the generalization issue that is | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 15 | gap in the many-to-one model between the training set BLEU and the development set BLEU,
which points on the generalization issue that is
not present in the many-to-many setting. We also
note that our many-to-one model is on average
0.75 BLEU behind the best many-to-one models
in Neubig and Hu (2018). We attribute this to the
fact that their models are fine-tuned using similarlanguage-regularization while our model is not.
We find an additional difference between the
results on the resource-scarce languages (Table 1) and the higher-resource languages (Table
2). Specifically, the bilingual baselines outperform the many-to-one models only in the higherresource setting. This makes sense as in the low-En-Az En-Be En-Gl En-Sk Avg.
# of examples 5.9k 4.5k 10k 61k 20.3k
baselines 2.16 2.47 3.26 5.8 3.42
one-to-many 5.06 10.72 26.59 24.52 16.72
many-to-many 3.9 7.24 23.78 21.83 14.19
En-Ar En-De En-He En-It Avg. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 16 | many-to-many 3.9 7.24 23.78 21.83 14.19
En-Ar En-De En-He En-It Avg.
# of examples 213k 167k 211k 203k 198.5k
baselines 12.95 23.31 23.66 30.33 22.56
one-to-many 16.67 30.54 27.62 35.89 27.68
many-to-many 14.25 27.95 24.16 33.26 24.9
Table 3: En !X test BLEU on the TED Talks corpus
resource setting the baselines have very few training examples to outperform the many-to-one models, while in the higher resource setting they have
access to more training data. This corroborates the
results of Gu et al. (2018) that showed the sensitivity of such models to similar low resource conditions and the improvements gained from using
many-to-one models (however with much fewer
language pairs).
Table 3 shows the results of our massively
multilingual models and bilingual baselines when
evaluated out-of-English. In this case we see an
opposite trend: the many-to-many model performs
worse than the one-to-many model by 2.53 BLEU
on average. While previous works (Wang et al., | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 17 | worse than the one-to-many model by 2.53 BLEU
on average. While previous works (Wang et al.,
2018; Sachan and Neubig, 2018) discuss the phenomena of quality degradation in English-to-many
settings, this shows that increasing the number of
source languages also causes additional degradation in a many-to-many model. This degradation
may be due to the English-centric setting: since
most of the translation directions the model is
trained on are into English, this leaves less capacity for the other target languages (while still performing better than the bilingual baselines on all 8
language pairs). We also note that in this case the
results are consistent among the higher and lower
resource pairs – the one-to-many model is better
than the many-to-many model, which outperforms
the bilingual baselines in all cases. This is unlike
the difference we saw in the X !En experiments
since here we do not have the multi-way-parallel
overfitting issue.
2.4 Discussion
From the above experiments we learn that NMT
models can scale to 59 languages in a lowresource, imbalanced, English-centric setting,
with the following observations: (1) massively
multilingual many-to-many models outperform | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 18 | models can scale to 59 languages in a lowresource, imbalanced, English-centric setting,
with the following observations: (1) massively
multilingual many-to-many models outperform
many-to-one and bilingual models with similar capacity and identical training conditions when averaged over 8 language pairs into English. We
attribute this improvement over the many-to-one
models to the multiple target language pairs which
may act as regularizers, especially in this lowresource multi-way-parallel setting that is prone to
memorization. (2) many-to-many models are inferior in performance when going out-of-English
in comparison to a one-to-many model. We attribute this to English being over-represented in
the English-centric many-to-many setting, where
it appears as a target language in 58 out of 116
trained directions, which may harm the performance on the rest of the target languages as the
model capacity is limited.3
It is important to stress the fact that we compared the different models under identical training
conditions and did not perform extensive hyperparameter tuning for each setting separately. However, we believe that such tuning may improve
performance even further, as the diversity in each
training batch is very different between the different settings. For example, while the baseline
model batches include only one language in the | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 19 | performance even further, as the diversity in each
training batch is very different between the different settings. For example, while the baseline
model batches include only one language in the
source and one language in the target, the manyto-many model includes 59 languages in each
side with a strong bias towards English. These
differences may require tailored hyper-parameter
choices for each settings (i.e. different batch sizes,
learning rate schedules, dropout rates etc.) which
would be interesting to explore in future work.
In the following experiments we investigate
whether these observations hold using (1) an even
larger set of languages, and (2) a much larger,
balanced training corpus that is not multi-wayparallel.
3 High-Resource Setting: 103 Languages
3.1 Experimental Setup
In this setting we scale the number of languages
and examples per language pair further when
training a single massively multilingual model.
Since we are not aware of a publicly available resource for this purpose, we construct an in-house
dataset. This dataset includes 102 language pairs
which we “mirror” to-and-from English, with up
to one million examples per language pair. This
results in 103 languages in total, and 204 translation directions which we train simultaneously. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 20 | to one million examples per language pair. This
results in 103 languages in total, and 204 translation directions which we train simultaneously.
3This issue may be alleviated by over-sampling the nonEnglish-target pairs, but we leave this for future work.More details about this dataset are available in Table 4, and Table 10 in the supplementary material
details all the languages in the dataset.4
Similarly to our previous experiments, we compare the massively multilingual models to bilingual baselines trained on the same data. We tokenize the data using an in-house tokenizer and then
apply joint subword segmentation to achieve an
open-vocabulary. In this setting we used a vocabulary of 64k subwords rather than 32k. Since the
dataset contains 24k unique characters, a 32k symbol vocabulary will consist of mostly characters,
thereby increasing the average sequence length.
Regarding the model, for these experiments we
use a larger Transformer model with 6 layers in
both the encoder and the decoder, model dimension set to 1024, hidden dimension size of 8192,
and 16 attention heads. This results in a model
with approximately 473.7M parameters.5Since
the model and data are much larger in this case,
we used a dropout rate of 0.1 for our multilingual
models and tuned it to 0.3 for our baseline models | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 21 | the model and data are much larger in this case,
we used a dropout rate of 0.1 for our multilingual
models and tuned it to 0.3 for our baseline models
as it improved the translation quality on the development set.
We evaluate our models on 10 languages from
different typological families: Semitic – Arabic
(Ar), Hebrew (He), Romance – Galician (Gl),
Italian (It), Romanian (Ro), Germanic – German
(De), Dutch (Nl), Slavic – Belarusian (Be), Slovak (Sk) and Turkic – Azerbaijani (Az) and Turkish (Tr). We evaluate both to-and-from English,
where each language pair is trained on up to one
million examples. As in the previous experiment,
we report test results from the model that performed best in terms of BLEU on the development
set.
4The average number of examples per language pair is
940k, as for 13 out of the 102 pairs we had less than one
million examples available.
5This is larger than the Transformer “Big” configuration,
which includes approximately 213M trained parameters.
# of language pairs 102
examples per pair
min 63,879
max 1,000,000
average 940,087
std. deviation 188,194 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 22 | # of language pairs 102
examples per pair
min 63,879
max 1,000,000
average 940,087
std. deviation 188,194
total # of examples 95,888,938
Table 4: Training set details for the 103 langauges corpus, X!En data.
Ar Az Be De He It Nl Ro Sk Tr Avg.
baselines 23.34 16.3 21.93 30.18 31.83 36.47 36.12 34.59 25.39 27.13 28.33
many-to-one 26.04 23.68 25.36 35.05 33.61 35.69 36.28 36.33 28.35 29.75 31.01
many-to-many 22.17 21.45 23.03 37.06 30.71 35.0 36.18 36.57 29.87 27.64 29.97
Table 5: X !En test BLEU on the 103-language corpus
Ar Az Be De He It Nl Ro Sk Tr Avg.
baselines 10.57 8.07 15.3 23.24 19.47 31.42 28.68 27.92 11.08 15.54 19.13
one-to-many 12.08 9.92 15.6 31.39 20.01 33 31.06 28.43 17.67 17.68 21.68 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 23 | one-to-many 12.08 9.92 15.6 31.39 20.01 33 31.06 28.43 17.67 17.68 21.68
many-to-many 10.57 9.84 14.3 28.48 17.91 30.39 29.67 26.23 18.15 15.58 20.11
Table 6: En !X test BLEU on the 103-language corpus
3.2 Results
Table 5 describes the results when translating into
English. First, we can see that both multilingual
models perform better than the baselines in terms
of average BLEU. This shows that massively multilingual many-to-many models can work well in
realistic settings with millions of training examples, 102 languages and 204 jointly trained directions to-and-from English. Looking more closely,
we note several different behaviors in comparison
to the low-resource experiments on the TED Talks
corpus. First, the many-to-one model here performs better than the many-to-many model. This
shows that the previous result was indeed due to
the pathologies of the low-resource dataset; when
the training data is large enough and not multiway-parallel there is no overfitting in the many-toone model, and it outperforms the many-to-many
model in most cases while they are trained identically. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 24 | model in most cases while they are trained identically.
One particular outlier in this case is German-toEnglish, where the many-to-one model is 2 BLEU
points below the many-to-many model. We examine the BLEU score of this language pair on its
dedicated German-English development set during training in the many-to-one model and find
that it highly fluctuates. We then measure the
performance on the test set for this language pair
by choosing the best checkpoint on the dedicated
German-English development set (instead of on
the mixed multilingual development set) and find
it to be 38.07, which is actually higher in 1 BLEU
than the best result of the many-to-many model.
This shows that while training many languages together, there is no “silver bullet”: some languages
may suffer from severe interference during training (i.e. a reduction of 3 BLEU in this case, from38.07 to 35.05) while other languages continue to
improve with more updates.
Table 6 describes the results when translating
out-of-English. Again, both of the massively multilingual models perform better than the baselines when averaged across the 10 evaluated language pairs, while handling up to 102 languages | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 25 | out-of-English. Again, both of the massively multilingual models perform better than the baselines when averaged across the 10 evaluated language pairs, while handling up to 102 languages
to-and-from English and 204 translation tasks simultaneously. In this case the results are similar to those we observed on the TED talks corpus, where the one-to-many model performs better
than the many-to-many model. Again, this advantage may be due to the one-to-many model handling a smaller number of tasks while not being
biased towards English in the target side like the
many-to-many model.
4 Analysis
The above results show that massively multilingual NMT is indeed possible in large scale settings
and can improve performance over strong bilingual baselines. However, it was shown in a somewhat extreme case with more than 100 languages
trained jointly, where we saw that in some cases
the joint training may harm the performance for
some language pairs (i.e. German-English above).
In the following analysis we would like to better understand the trade-off between the number
of languages involved and the translation accuracy while keeping the model capacity and training configuration fixed.
4.1 Multilinguality & Supervised
Performance
We first study the effect of varying the number of languages on the translation accuracy in | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 26 | 4.1 Multilinguality & Supervised
Performance
We first study the effect of varying the number of languages on the translation accuracy in
a supervised setting, where we focus on manyAr-En En-Ar Fr-En En-Fr Ru-En En-Ru Uk-En En-Uk Avg.
5-to-5 23.87 12.42 38.99 37.3 29.07 24.86 26.17 16.48 26.14
25-to-25 23.43 11.77 38.87 36.79 29.36 23.24 25.81 17.17 25.8
50-to-50 23.7 11.65 37.81 35.83 29.22 21.95 26.02 15.32 25.18
75-to-75 22.23 10.69 37.97 34.35 28.55 20.7 25.89 14.59 24.37
103-to-103 21.16 10.25 35.91 34.42 27.25 19.9 24.53 13.89 23.41
Table 7: Supervised performance while varying the number of languages involved
Ar-Fr Fr-Ar Ru-Uk Uk-Ru Avg.
5-to-5 1.66 4.49 3.7 3.02 3.21 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 27 | Ar-Fr Fr-Ar Ru-Uk Uk-Ru Avg.
5-to-5 1.66 4.49 3.7 3.02 3.21
25-to-25 1.83 5.52 16.67 4.31 7.08
50-to-50 4.34 4.72 15.14 20.23 11.1
75-to-75 1.85 4.26 11.2 15.88 8.3
103-to-103 2.87 3.05 12.3 18.49 9.17
Table 8: Zero-Shot performance while varying the
number of languages involved
to-many models. We create four subsets of the
in-house dataset by sub-sampling it to a different number of languages in each subset. In
this way we create four additional English-centric
datasets, containing 5, 25, 50 and 75 languages
each to-and-from English. We make sure that
each subset contains all the languages from the
next smaller subsets – i.e. the 25 language subset contains the 5 language subset, the 50 language subset contains the 25 language subset and
so on. We train a similar-capacity large Transformer model (with 473.7M parameters) on each
of these subsets and measure the performance for
each model on the 8 supervised language pairs | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 28 | of these subsets and measure the performance for
each model on the 8 supervised language pairs
from the smallest subset – fArabic, French, Russian, Ukrainian g$English. In this way we can
analyze to what extent adding more languages improves or harms translation quality while keeping
the model capacity fixed, testing the capacity vs.
accuracy “saturation point”.
Table 7 shows the results of this experiment,
reporting the test results for the models that performed best on the multilingual development set.
We can see that in most cases the best results
are obtained using the 5-to-5 model, showing that
there is indeed a trade off between the number
of languages and translation accuracy when using a fixed model capacity and the same training setup. One may expect that the gaps between
the different models should become smaller and
even close with more updates, as the models with
more languages see less examples per language
in each batch, thus requiring more updates to improve in terms of BLEU. However, in our setting
these gaps did not close even after the models converged, leaving 2.73 average BLEU difference be-tween the 5-to-5 and the 103-to-103 model.
4.2 Multilinguality & Zero-Shot
Performance
We then study the effect of the number of languages on zero-shot translation accuracy. Since | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 29 | 4.2 Multilinguality & Zero-Shot
Performance
We then study the effect of the number of languages on zero-shot translation accuracy. Since
we find zero-shot accuracy as an interesting measure for model generalization, we hypothesize that
by adding more languages, the model is forced to
create a more generalized representation to better utilize its capacity, which may improve zeroshot performance. We choose four language pairs
for this purpose: Arabic $French which are distant languages, and Ukrainian $Russian which are
similar. Table 8 shows the results of our models
on these language pairs. For Arabic $French the
BLEU scores are very low in all cases, with the
50-to-50 and 25-to-25 models being slightly better than rest on Ar-Fr and Fr-Ar respectively. On
Russian $Ukrainian we see clear improvements
when increasing the number of languages to more
than five.
Figure 2 further illustrates this, showing the better generalization performance of the massively
multilingual models under this zero-shot setting.
While the zero-shot performance in this case is
low and unstable for the 5-to-5 and 25-to-25 modupdate 5-to-5 25-to-25 50-to-50 75-to-75 103-to-103
0 0 0 0 0 0 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 30 | 0 0 0 0 0 0
25000 3.360397369 7.345648855 7.028211653 6.687645614 6.629930437
50000 2.003555186 10.08476391 11.66040972 11.46485135 11.34905592
100000 2.383616194 9.54657495 15.44517726 14.72926438 13.17522973
150000 2.588021383 7.121089101 14.5408541 15.54533243 15.89359492
200000 2.589718811 10.99432111 15.82096368 17.38970876 16.26121253
250000 2.854427323 14.00393397 16.53215885 15.97282737 15.3215304
300000 2.823847346 6.934611499 16.78672731 16.52613729 16.16107672
350000 2.950227261 4.779103771 16.73545986 16.12752229 17.03165472
400000 4.301280901 6.631205231 16.16190225 17.38892049 17.55904406 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 31 | 400000 4.301280901 6.631205231 16.16190225 17.38892049 17.55904406
450000 3.882381693 6.804813445 18.20554733 17.48778224 17.9339543
500000 3.45445089 6.358428299 17.32598543 16.71108902 15.97367823
550000 3.18 6.38 18.65 18.28 17.4
600000 2.86 9.5 18.46 14.92 17.12
650000 2.55 12.2 18.98 15.68 16.19
700000 2.98 8.44 20.16 15.4 18.52
Figure 2: Zero-shot BLEU during training for Ukranian to Russian
els, it is much better for the 50-to-50, 75-to-75 and
103-to-103 models. Given these results we can
say that the balance between capacity and generalization here favors the mid range 50-to-50 model,
even when using models with more than 473M
trained parameters. This may hint at the necessity of even larger models for such settings, which
is a challenging avenue for future work. We also
note that our 103 language corpus includes up to | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 32 | trained parameters. This may hint at the necessity of even larger models for such settings, which
is a challenging avenue for future work. We also
note that our 103 language corpus includes up to
one million examples per language pair – while in
real-world MT deployments, systems are trained
on much more examples per pair. This again emphasizes the need for better techniques for training
such massively multilingual models as we may already be hitting the capacity barrier in our setting.
5 Related Work
Dong et al. (2015) extended the NMT model of
Bahdanau et al. (2014) to one-to-many translation
(from English into 4 languages) by adding a dedicated decoder per target language, showing improvements over strong single-pair baselines. Firat et al. (2016a,b) proposed many-to-many models (with up to 6 languages) by using separate encoders and decoders per language while sharing
the attention mechanism. They also introduced
the notion of zero-resource translation, where they
use synthetic training data generated through pivoting to train translation directions without available training data. Ha et al. (2016) and Johnson
et al. (2017) proposed to use a shared encoderdecoder-attention model for many-to-many translation (with up to 7 languages in the latter). In
order to determine the target language in such | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 33 | order to determine the target language in such
scenarios they proposed adding dedicated targetlanguage symbols to the source. This method enabled zero-shot translation, showing the ability of
the model to generalize to unseen pairs.
Recent works propose different methods for parameter sharing between language pairs in multilingual NMT. Blackwood et al. (2018) propose
sharing all parameters but the attention mechanism
and show improvements over sharing all parameters. Sachan and Neubig (2018) explore sharing various components in self-attentional (Transformer) models. Lu et al. (2018) add a shared “interlingua” layer while using separate encoders and
decoders. Zaremoodi et al. (2018) utilize recurrent
units with multiple blocks together with a trainable
routing network. Platanios et al. (2018) propose
to share the entire network, while using a contex-tual parameter generator that learns to generate the
parameters of the system given the desired source
and target languages. Gu et al. (2018) propose
a “Universal Language Representation” layer together with a Mixture-of-Language-Experts component to improve a many-to-one model from 5
languages into English.
While the mentioned studies provide valuable
contributions to improving multilingual models,
they apply their models on only up to 7 languages | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 34 | languages into English.
While the mentioned studies provide valuable
contributions to improving multilingual models,
they apply their models on only up to 7 languages
(Johnson et al., 2017) and 20 trained directions
(Cettolo et al., 2017) in a single model, whereas
we focus on scaling NMT to much larger numbers of languages and trained directions. Regarding massively multilingual models, Neubig and
Hu (2018) explored methods for rapid adaptation
of NMT to new languages by training multilingual models on the 59-language TED Talks corpus and fine-tuning them using data from the new
languages. While modeling significantly more
languages than previous studies, they only train
many-to-one models, which we show are inferior
in comparison to our proposed massively multilingual many-to-many models when evaluated into
English on this dataset.
Tiedemann (2018) trained an English-centric
many-to-many model on translations of the bible
including 927 languages. While this work pointed
to an interesting phenomena in the latent space
learned by the model where it clusters representations of typologically-similar languages together, it did not include any evaluation of the
produced translations. Similarly, Malaviya et al.
(2017) trained a many-to-English system including 1017 languages from bible translations, and | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 35 | produced translations. Similarly, Malaviya et al.
(2017) trained a many-to-English system including 1017 languages from bible translations, and
used it to infer typological features for the different languages (without evaluating the translation quality). In another relevant work, Artetxe
and Schwenk (2018) trained an NMT model on
93 languages and used the learned representations
to perform cross-lingual transfer learning. Again,
they did not report the performance of the translation model learned in that massively multilingual
setting.
6 Conclusions and Future Work
We showed that NMT models can successfully
scale to 102 languages to-and-from English with
204 trained directions and up to one million examples per direction. Such models improve the
translation quality over similar single-pair baselines when evaluated to and from English by more
than 2 BLEU when averaged over 10 diverse language pairs in each case. We show a similar result on the low-resource TED Talks corpus with 59
languages and 116 trained directions. We analyze
the trade-offs between translation quality and the
number of languages involved, pointing on capacity bottlenecks even with very large models and
showing that massively multilingual models can
generalize better to zero-shot settings.
We hope this work will encourage future research on massively multilingual NMT, enabling | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 36 | showing that massively multilingual models can
generalize better to zero-shot settings.
We hope this work will encourage future research on massively multilingual NMT, enabling
easier support for systems that can serve more people around the globe. There are many possible avenues for future work, including semi-supervised
learning in such settings, exploring ways to reduce the performance degradation when increasing the number of languages, or using such models
for multilingual transfer learning (McCann et al.,
2017; Eriguchi et al., 2018; Artetxe and Schwenk,
2018). Understanding and improving zero-shot
performance in such scenarios is also a promising
direction for future work.
Acknowledgments
We would like to thank the Google Brain and
Google Translate teams for their useful inputs and
discussions. We would also like to thank the entire
Lingvo development team for their foundational
contributions to this project.
References
Mikel Artetxe and Holger Schwenk. 2018. Massively multilingual sentence embeddings for zeroshot cross-lingual transfer and beyond. arXiv
preprint arXiv:1812.10464 .
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly
learning to align and translate. arXiv preprint | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 37 | Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly
learning to align and translate. arXiv preprint
arXiv:1409.0473 .
Graeme Blackwood, Miguel Ballesteros, and Todd
Ward. 2018. Multilingual neural machine translation with task-specific attention. In Proceedings of
the 27th International Conference on Computational
Linguistics , Santa Fe, New Mexico, USA. Association for Computational Linguistics.
Ondˇrej Bojar, Rajen Chatterjee, Christian Federmann,
et al. 2017. Findings of the 2017 conference on machine translation (wmt17). In Proceedings of the
Second Conference on Machine Translation .Ondˇrej Bojar, Rajen Chatterjee, Christian Federmann,
et al. 2016. Findings of the 2016 conference on machine translation. In ACL 2016 FIRST CONFERENCE ON MACHINE TRANSLATION (WMT16) ,
pages 131–198.
Ondˇrej Bojar, Christian Federmann, Mark Fishel,
Yvette Graham, Barry Haddow, Philipp Koehn, and | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 38 | Ondˇrej Bojar, Christian Federmann, Mark Fishel,
Yvette Graham, Barry Haddow, Philipp Koehn, and
Christof Monz. 2018. Findings of the 2018 conference on machine translation (wmt18). In Proceedings of the Third Conference on Machine Translation: Shared Task Papers .
Mauro Cettolo, Federico Marcello, Bentivogli Luisa,
Niehues Jan, St ¨uker Sebastian, Sudoh Katsuitho,
Yoshino Koichiro, and Federmann Christian. 2017.
Overview of the iwslt 2017 evaluation campaign. In
International Workshop on Spoken Language Translation .
Daxiang Dong, Hua Wu, Wei He, Dianhai Yu, and
Haifeng Wang. 2015. Multi-task learning for multiple language translation. In Proceedings of the
53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint
Conference on Natural Language Processing .
Akiko Eriguchi, Melvin Johnson, Orhan Firat, Hideto
Kazawa, and Wolfgang Macherey. 2018. Zeroshot cross-lingual classification using multilingual neural machine translation. arXiv preprint
arXiv:1809.04686 . | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 39 | arXiv:1809.04686 .
Orhan Firat, Kyunghyun Cho, and Yoshua Bengio.
2016a. Multi-way, multilingual neural machine
translation with a shared attention mechanism. In
Proceedings of the 2016 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies .
Orhan Firat, Baskaran Sankaran, Yaser Al-Onaizan,
Fatos T Yarman Vural, and Kyunghyun Cho. 2016b.
Zero-resource translation with multi-lingual neural
machine translation. In Proceedings of the 2016
Conference on Empirical Methods in Natural Language Processing .
Jiatao Gu, Hany Hassan, Jacob Devlin, and Victor O.K.
Li. 2018. Universal neural machine translation for
extremely low resource languages. In Proceedings of the 2018 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies .
Thanh-Le Ha, Jan Niehues, and Alexander Waibel.
2016. Toward multilingual neural machine translation with universal encoder and decoder. arXiv
preprint arXiv:1611.04798 .
Hany Hassan, Anthony Aue, Chang Chen, et al. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 40 | preprint arXiv:1611.04798 .
Hany Hassan, Anthony Aue, Chang Chen, et al.
2018. Achieving human parity on automatic chinese to english news translation. arXiv preprint
arXiv:1803.05567 .
Melvin Johnson, Mike Schuster, Quoc V Le, et al.
2017. Google’s multilingual neural machine translation system: Enabling zero-shot translation. Transactions of the Association of Computational Linguistics.
Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent continuous translation models. In Proceedings
of the 2013 Conference on Empirical Methods in
Natural Language Processing , Seattle, Washington,
USA.
Surafel Melaku Lakew, Mauro Cettolo, and Marcello
Federico. 2018. A comparison of transformer and
recurrent neural networks on multilingual neural
machine translation. In Proceedings of the 27th International Conference on Computational Linguistics, Santa Fe, New Mexico, USA.
Yichao Lu, Phillip Keung, Faisal Ladhak, Vikas Bhardwaj, Shaonan Zhang, and Jason Sun. 2018. A neural interlingua for multilingual machine translation.
InProceedings of the Third Conference on Machine
Translation: Research Papers , Belgium, Brussels. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 41 | InProceedings of the Third Conference on Machine
Translation: Research Papers , Belgium, Brussels.
Chaitanya Malaviya, Graham Neubig, and Patrick Littell. 2017. Learning language representations for typology prediction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language
Processing .
Bryan McCann, James Bradbury, Caiming Xiong, and
Richard Socher. 2017. Learned in translation: Contextualized word vectors. In Advances in Neural Information Processing Systems .
Graham Neubig and Junjie Hu. 2018. Rapid adaptation
of neural machine translation to new languages. In
Proceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing .
Toan Q. Nguyen and David Chiang. 2017. Transfer
learning across low-resource, related languages for
neural machine translation. In Proc. IJCNLP .
Myle Ott, Sergey Edunov, David Grangier, and
Michael Auli. 2018. Scaling neural machine translation. In Proceedings of the Third Conference on
Machine Translation: Research Papers .
Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of
40th Annual Meeting of the Association for Computational Linguistics , Philadelphia, Pennsylvania,
USA. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 42 | 40th Annual Meeting of the Association for Computational Linguistics , Philadelphia, Pennsylvania,
USA.
Emmanouil Antonios Platanios, Mrinmaya Sachan,
Graham Neubig, and Tom Mitchell. 2018. Contextual parameter generation for universal neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language
Processing .Ye Qi, Sachan Devendra, Felix Matthieu, Padmanabhan Sarguna, and Neubig Graham. 2018. When
and why are pre-trained word embeddings useful for
neural machine translation. In HLT-NAACL .
Devendra Sachan and Graham Neubig. 2018. Parameter sharing methods for multilingual self-attentional
translation models. In Proceedings of the Third
Conference on Machine Translation: Research Papers, Belgium, Brussels.
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2016. Neural machine translation of rare words with
subword units. In Proceedings of the 54th Annual
Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , volume 1, pages
1715–1725.
Jonathan Shen, Patrick Nguyen, Yonghui Wu, Zhifeng
Chen, et al. 2019. Lingvo: a modular and scalable
framework for sequence-to-sequence modeling. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 43 | Jonathan Shen, Patrick Nguyen, Yonghui Wu, Zhifeng
Chen, et al. 2019. Lingvo: a modular and scalable
framework for sequence-to-sequence modeling.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
Sequence to sequence learning with neural networks. In Advances in neural information processing systems , pages 3104–3112.
J¨org Tiedemann. 2018. Emerging language spaces
learned from massively multilingual corpora. arXiv
preprint arXiv:1802.00273 .
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. In Advances in Neural Information Processing Systems , pages 5998–6008.
Xinyi Wang, Hieu Pham, Philip Arthur, and Graham
Neubig. 2019. Multilingual neural machine translation with soft decoupled encoding. In International
Conference on Learning Representations .
Yining Wang, Jiajun Zhang, Feifei Zhai, Jingfang Xu, | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 44 | Conference on Learning Representations .
Yining Wang, Jiajun Zhang, Feifei Zhai, Jingfang Xu,
and Chengqing Zong. 2018. Three strategies to improve one-to-many multilingual translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , Brussels, Belgium.
Yonghui Wu, Mike Schuster, Zhifeng Chen, et al. 2016.
Google’s neural machine translation system: Bridging the gap between human and machine translation.
arXiv preprint arXiv:1609.08144 .
Poorya Zaremoodi, Wray Buntine, and Gholamreza
Haffari. 2018. Adaptive knowledge sharing in
multi-task learning: Improving low-resource neural
machine translation. In Proceedings of the 56th Annual Meeting of the Association for Computational
Linguistics , Melbourne, Australia.
Barret Zoph, Deniz Yuret, Jonathan May, and Kevin
Knight. 2016. Transfer learning for low-resource
neural machine translation. In Proceedings of the
2016 Conference on Empirical Methods in Natural
Language Processing .
A Supplementary Material
Language Train set size
Arabic 214111
Hebrew 211819
Russian 208458
Korean 205640
Italian 204503
Japanese 204090
Chinese-Taiwan 202646 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 45 | Arabic 214111
Hebrew 211819
Russian 208458
Korean 205640
Italian 204503
Japanese 204090
Chinese-Taiwan 202646
Chinese-China 199855
Spanish 196026
French 192304
Portuguese-Brazil 184755
Dutch 183767
Turkish 182470
Romanian 180484
Polish 176169
Bulgarian 174444
Vietnamese 171995
German 167888
Persian 150965
Hungarian 147219
Serbian 136898
Greek 134327
Croatian 122091
Ukrainian 108495
Czech 103093
Thai 98064
Indonesian 87406
Slovak 61470
Swedish 56647
Portuguese 51785
Danish 44940
Albanian 44525
Lithuanian 41919
Macedonian 25335
Finnish 24222
Burmese 21497
Armenian 21360
French-Canadian 19870
Slovenian 19831
Hindi 18798
Norwegian 15825
Georgian 13193
Estonian 10738
Kurdish 10371
Galician 10017
Marathi 9840
Mongolian 7607
Esperanto 6535
Tamil 6224
Urdu 5977
Azerbaijani 5946 | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 46 | Marathi 9840
Mongolian 7607
Esperanto 6535
Tamil 6224
Urdu 5977
Azerbaijani 5946
Bosnian 5664
Chinese 5534
Malay 5220
Basque 5182
Bengali 4649
Belarusian 4509
Kazakh 3317
Table 9: Language pairs in the TED talks dataset (58
languages, paired with English) with the train-set size
for each pair.Languages
Afrikaans Laothian
Albanian Latin
Amharic Latvian
Arabic Lithuanian
Armenian Luxembourgish*
Azerbaijani Macedonian
Basque Malagasy
Belarusian Malay
Bengali Malayalam
Bosnian Maltese
Bulgarian Maori
Burmese Marathi
Catalan Mongolian
Cebuano Nepali
Chichewa* Norwegian
Chinese Pashto
Corsican* Persian
Croatian Polish
Czech Portuguese
Danish Punjabi
Dutch Romanian
Esperanto Russian
Estonian Samoan*
Finnish Scots Gaelic*
French Serbian
Frisian Sesotho
Galician Shona*
Georgian Sindhi*
German Sinhalese
Greek Slovak
Gujarati Slovenian
Haitian Creole Somali | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1903.00089 | 47 | Galician Shona*
Georgian Sindhi*
German Sinhalese
Greek Slovak
Gujarati Slovenian
Haitian Creole Somali
Hausa* Spanish
Hawaiian* Sundanese
Hebrew Swahili
Hindi Swedish
Hmong* Tagalog
Hungarian Tajik*
Icelandic Tamil
Igbo Telugu
Indonesian Thai
Irish Turkish
Italian Ukrainian
Japanese Urdu
Javanese Uzbek
Kannada Vietnamese
Kazakh Welsh
Khmer Xhosa
Korean Yiddish
Kurdish Yoruba*
Kyrgyz Zulu
Table 10: Language pairs in the in-house dataset (102
languages, paired with English). For languages marked
with * we had less than 1M examples, while for the rest
we used exactly 1M. | 1903.00089 | Massively Multilingual Neural Machine Translation | Multilingual neural machine translation (NMT) enables training a single model
that supports translation from multiple source languages into multiple target
languages. In this paper, we push the limits of multilingual NMT in terms of
number of languages being used. We perform extensive experiments in training
massively multilingual NMT models, translating up to 102 languages to and from
English within a single model. We explore different setups for training such
models and analyze the trade-offs between translation quality and various
modeling decisions. We report results on the publicly available TED talks
multilingual corpus where we show that massively multilingual many-to-many
models are effective in low resource settings, outperforming the previous
state-of-the-art while supporting up to 59 languages. Our experiments on a
large-scale dataset with 102 languages to and from English and up to one
million examples per direction also show promising results, surpassing strong
bilingual baselines and encouraging future work on massively multilingual NMT. | http://arxiv.org/pdf/1903.00089 | [
"Roee Aharoni",
"Melvin Johnson",
"Orhan Firat"
] | [
"cs.CL"
] | Accepted as a long paper in NAACL 2019 | null | cs.CL | 20190228 | 20190702 | [] |
1904.09925 | 0 | Attention Augmented Convolutional Networks
Irwan Bello Barret Zoph Ashish Vaswani Jonathon Shlens
Quoc V . Le
Google Brain
fibello,barretzoph,avaswani,shlens,qvl [email protected]
Abstract
Convolutional networks have been the paradigm of
choice in many computer vision applications. The convolution operation however has a significant weakness in that it
only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged
as a recent advance to capture long range interactions, but
has mostly been applied to sequence modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to convolutions. We introduce a novel two-dimensional
relative self-attention mechanism that proves competitive
in replacing convolutions as a stand-alone computational
primitive for image classification. We find in control experiments that the best results are obtained when combining
both convolutions and self-attention. We therefore propose
to augment convolutional operators with this self-attention
mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments show that Attention Augmentation leads | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 1 | mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments show that Attention Augmentation leads
to consistent improvements in image classification on ImageNet and object detection on COCO across many different models and scales, including ResNets and a stateof-the art mobile constrained network, while keeping the
number of parameters similar. In particular, our method
achieves a 1:3%top-1 accuracy improvement on ImageNet
classification over a ResNet50 baseline and outperforms
other attention mechanisms for images such as Squeezeand-Excitation [17]. It also achieves an improvement of
1.4 mAP in COCO Object Detection on top of a RetinaNet
baseline.
1. Introduction
Convolutional Neural Networks have enjoyed tremendous success in many computer vision applications, especially in image classification [24, 23]. The design of the
convolutional layer imposes 1) locality via a limited receptive field and 2) translation equivariance via weight sharing.
20 30 40 50 60 70
number of parameters (millions)72737475767778798081top-1 accuracy
73.6%76.4%77.9%78.4% | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 2 | number of parameters (millions)72737475767778798081top-1 accuracy
73.6%76.4%77.9%78.4%
74.3%77.5%78.4%78.9%
74.7%77.7%78.7%79.1%
ResNet
SE-ResNet
AA-ResNet (ours)Figure 1. Attention Augmentation systematically improves image classification across a large variety of networks of different
scales. ImageNet classification accuracy [9] versus the number of
parameters for baseline models (ResNet) [14], models augmented
with channel-wise attention (SE-ResNet) [17] and our proposed
architecture (AA-ResNet).
Both these properties prove to be crucial inductive biases
when designing models that operate over images. However,
the local nature of the convolutional kernel prevents it from
capturing global contexts in an image, often necessary for
better recognition of objects in images [33].
Self-attention [43], on the other hand, has emerged as a
recent advance to capture long range interactions, but has
mostly been applied to sequence modeling and generative
modeling tasks. The key idea behind self-attention is to | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 3 | recent advance to capture long range interactions, but has
mostly been applied to sequence modeling and generative
modeling tasks. The key idea behind self-attention is to
produce a weighted average of values computed from hidden units. Unlike the pooling or the convolutional operator,
the weights used in the weighted average operation are produced dynamically via a similarity function between hidden units. As a result, the interaction between input signals
depends on the signals themselves rather than being predetermined by their relative location like in convolutions. In
particular, this allows self-attention to capture long rangearXiv:1904.09925v5 [cs.CV] 9 Sep 2020
Input Attention maps Weighted average of the values
Standard convolutionOutput
W
H
Nh=2
Values²Head Head¹
Figure 2. Attention-augmented convolution : For each spatial location (h, w), Nhattention maps over the image are computed from
queries and keys. These attention maps are used to compute Nhweighted averages of the values V . The results are then concatenated,
reshaped to match the original volume’s spatial dimensions and mixed with a pointwise convolution. Multi-head attention is applied in
parallel to a standard convolution operation and the outputs are concatenated.
interactions without increasing the number of parameters. | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 4 | parallel to a standard convolution operation and the outputs are concatenated.
interactions without increasing the number of parameters.
In this paper, we consider the use of self-attention for
discriminative visual tasks as an alternative to convolutions. We develop a novel two-dimensional relative selfattention mechanism [37] that maintains translation equivariance while being infused with relative position information, making it well suited for images. Our self-attention
formulation proves competitive for replacing convolutions
entirely, however we find in control experiments that the
best results are obtained when combining both. We therefore do not completely abandon the idea of convolutions,
but instead propose to augment convolutions with this selfattention mechanism. This is achieved by concatenating
convolutional feature maps, which enforce locality, to selfattentional feature maps capable of modeling longer range
dependencies (see Figure 2).
We test our method on the CIFAR-100 and ImageNet
classification [22, 9] and the COCO object detection [27]
tasks, across a wide range of architectures at different computational budgets, including a state-of-the art resource
constrained architecture [42]. Attention Augmentation
yields systematic improvements with minimal additional | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 5 | constrained architecture [42]. Attention Augmentation
yields systematic improvements with minimal additional
computational burden and notably outperforms the popular Squeeze-and-Excitation [17] channelwise attention approach in all experiments. In particular, Attention Augmentation achieves a 1.3% top-1 accuracy ImageNet on top of
a ResNet50 baseline and 1.4 mAP increase in COCO object detection on top of a RetinaNet baseline. Suprisingly,
experiments also reveal that fully self-attentional models,
a special case of Attention Augmentation, only perform
slightly worse than their fully convolutional counterparts on
ImageNet, indicating that self-attention is a powerful standalone computational primitive for image classification.2. Related Work
2.1. Convolutional networks
Modern computer vision has been built on powerful image featurizers learned on image classification tasks such
as CIFAR-10 [22] and ImageNet [9]. These datasets have
been used as benchmarks for delineating better image featurizations and network architectures across a broad range
of tasks [21]. For example, improving the “backbone” network typically leads to improvements in object detection
[19] and image segmentation [6]. These observations have
inspired the research and design of new architectures, which
are typically derived from the composition of convolution | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 6 | [19] and image segmentation [6]. These observations have
inspired the research and design of new architectures, which
are typically derived from the composition of convolution
operations across an array of spatial scales and skip connections [23, 41, 39, 40, 14, 47, 13]. Indeed, automated
search strategies for designing architectures based on convolutional primitives result in state-of-the-art accuracy on
large-scale image classification tasks that translate across a
range of tasks [55, 21].
2.2. Attention mechanisms in networks
Attention has enjoyed widespread adoption as a computational module for modeling sequences because of its
ability to capture long distance interactions [2, 44, 4, 3].
Most notably, Bahdanau et al. [2] first proposed to combine attention with a Recurrent Neural Network [15] for
alignment in Machine Translation. Attention was further
extended by Vaswani et al. [43], where the self-attentional
Transformer architecture achieved state-of-the-art results in
Machine Translation. Using self-attention in cooperation
with convolutions is a theme shared by recent work in Natural Language Processing [49] and Reinforcement Learning [52]. For example, the QANet [50] and Evolved Transformer [38] architectures alternate between self-attention | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 7 | layers and convolution layers for Question Answering applications and Machine Translation respectively. Additionally, multiple attention mechanisms have been proposed for visual tasks to address the weaknesses of convolutions [17, 16, 7, 46, 45, 53]. For instance, Squeezeand-Excitation [17] and Gather-Excite [16] reweigh feature
channels using signals aggregated from entire feature maps,
while BAM [31] and CBAM [46] refine convolutional features independently in the channel and spatial dimensions.
In non-local neural networks [45], improvements are shown
in video classification and object detection via the additive use of a few non-local residual blocks that employ
self-attention in convolutional architectures. However, nonlocal blocks are only added to the architecture after ImageNet pretraining and are initialized in such a way that they
do not break pretraining.
In contrast, our attention augmented networks do not rely
on pretraining of their fully convolutional counterparts and
employ self-attention along the entire architecture. The use
of multi-head attention allows the model to attend jointly | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 8 | on pretraining of their fully convolutional counterparts and
employ self-attention along the entire architecture. The use
of multi-head attention allows the model to attend jointly
to both spatial and feature subspaces. Additionally, we enhance the representational power of self-attention over images by extending relative self-attention [37, 18] to two dimensional inputs allowing us to model translation equivariance in a principled way. Finally our method produces additional feature maps, rather than recalibrating convolutional
features via addition [45, 53] or gating [17, 16, 31, 46]. This
property allows us to flexibly adjust the fraction of attentional channels and consider a spectrum of architectures,
ranging from fully convolutional to fully attentional models.
3. Methods
We now formally describe our proposed Attention Augmentation method. We use the following naming conventions:H,WandFinrefer to the height, width and number
of input filters of an activation map. Nh,dvanddkrespectively refer the number of heads, the depth of values and the
depth of queries and keys in multihead-attention (MHA).
We further assume that Nhdividesdvanddkevenly and
denotedh
vanddh
kthe depth of values and queries/keys per
attention head. | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 9 | We further assume that Nhdividesdvanddkevenly and
denotedh
vanddh
kthe depth of values and queries/keys per
attention head.
3.1. Self-attention over images
Given an input tensor of shape (H;W;F in),1we flatten
it to a matrix X2RHWFinand perform multihead attention as proposed in the Transformer architecture [43]. The
output of the self-attention mechanism for a single head h
1We omit the batch dimension for simplicity.can be formulated as:
Oh=Softmax0
@(XW q)(XW k)T
q
dh
k1
A(XW v)(1)
whereWq,Wk2RFindh
kandWv2RFindh
vare learned
linear transformations that map the input Xto queriesQ=
XW q, keysK=XW kand valuesV=XW v. The outputs
of all heads are then concatenated and projected again as
follows:
MHA (X) =Concath
O1;:::;O Nhi
WO(2)
whereWO2Rdvdvis a learned linear transformation. | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 10 | O1;:::;O Nhi
WO(2)
whereWO2Rdvdvis a learned linear transformation.
MHA (X)is then reshaped into a tensor of shape (H;W;d v)
to match the original spatial dimensions. We note that
multi-head attention incurs a complexity of O((HW)2dk)
and a memory cost of O((HW)2Nh)as it requires to store
attention maps for each head.
3.1.1 Two-dimensional Positional Embeddings
Without explicit information about positions, self-attention
ispermutation equivariant :
MHA ((X)) =(MHA (X))
for any permutation of the pixel locations, making it ineffective for modeling highly structured data such as images. Multiple positional encodings that augment activation
maps with explicit spatial information have been proposed
to alleviate related issues. In particular, the Image Transformer [32] extends the sinusoidal waves first introduced in
the original Transformer [43] to 2 dimensional inputs and
CoordConv [29] concatenates positional channels to an activation map. | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 11 | the original Transformer [43] to 2 dimensional inputs and
CoordConv [29] concatenates positional channels to an activation map.
However these encodings did not help in our experiments on image classification and object detection (see Section 4.5). We hypothesize that this is because such positional encodings, while not permutation equivariant, do not
satisfy translation equivariance , which is a desirable property when dealing with images. As a solution, we propose
to extend the use of relative position encodings [37] to two
dimensions and present a memory efficient implementation
based on the Music Transformer [18].
Relative positional embeddings: Introduced in [37] for
the purpose of language modeling, relative self-attention
augments self-attention with relative position embeddings
and enables translation equivariance while preventing permutation equivariance. We implement two-dimensional relative self-attention by independently adding relative height
information and relative width information. The attention
logit for how much pixel i= (ix;iy)attends to pixel
j= (jx;jy)is computed as:
li;j=qT
iq
dh
k(kj+rW
jx ix+rH
jy iy) (3) | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 12 | li;j=qT
iq
dh
k(kj+rW
jx ix+rH
jy iy) (3)
whereqiis the query vector for pixel i(the i-th row of Q),
kjis the key vector for pixel j(the j-th row of K) andrW
jx ix
andrH
jy iyare learned embeddings for relative width jx ix
and relative height jy iy, respectively. The output of head
hnow becomes:
Oh=Softmax0
@QKT+Srel
H+Srel
Wq
dh
k1
AV (4)
whereSrel
H;Srel
W2RHWHWare matrices of relative position logits along height and width dimensions that satisfy
Srel
H[i;j] =qT
irH
jy iyandSrel
W[i;j] =qT
irW
jx ix.
The relative attention algorithm in [37] explicitly
stores all relative embeddings rijin a tensor of shape
(HW;HW;dh
k), thus incurring an additional memory cost
ofO((HW)2dh
k). This compares to O((HW)2Nh)for the | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 13 | k), thus incurring an additional memory cost
ofO((HW)2dh
k). This compares to O((HW)2Nh)for the
position-unaware version self-attention that does not use
position encodings. As we typically have Nh<dh
k, such an
implementation can prove extremely prohibitive and restrict
the number of images that can fit in a minibatch. Instead, we
extend the memory efficient relative masked attention algorithm presented in [18] to unmasked relative self-attention
over 2 dimensional inputs. Our implementation has a memory cost ofO(HWdh
k). We leave the Tensorflow code of
the algorithm in the Appendix.
The relative positional embeeddings rHandrWare
learned and shared across heads but not layers. For each
layer, we add (2(H+W) 2)dh
kparameters to model relative distances along height and width.
3.2. Attention Augmented Convolution
Multiple previously proposed attention mechanisms over
images [17, 16, 31, 46] suggest that the convolution operator is limited by its locality and lack of understanding
of global contexts. These methods capture long-range dependencies by recalibrating convolutional feature maps. In | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 14 | of global contexts. These methods capture long-range dependencies by recalibrating convolutional feature maps. In
particular, Squeeze-and-Excitation (SE) [17] and GatherExcite (GE) [16] perform channelwise reweighing while
BAM [31] and CBAM [46] reweigh both channels and
spatial positions independently . In contrast to these approaches, we 1) use an attention mechanism that can attend
jointly to spatial and feature subspaces (each head corresponding to a feature subspace) and 2) introduce additional
feature maps rather than refining them. Figure 2 summarizes our proposed augmented convolution.Concatenating convolutional and attentional feature
maps: Formally, consider an original convolution operator with kernel size k,Fininput filters and Foutoutput
filters. The corresponding attention augmented convolution
can be written as
AAConv (X) =Concath
Conv (X);MHA (X)i
:
We denote =dv
Foutthe ratio of attentional channels to
number of original output filters and =dk
Foutthe ratio of
key depth to number of original output filters. Similarly to | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 15 | number of original output filters and =dk
Foutthe ratio of
key depth to number of original output filters. Similarly to
the convolution, the proposed attention augmented convolution 1) is equivariant to translation and 2) can readily operate on inputs of different spatial dimensions. We include
Tensorflow code for the proposed attention augmented convolution in the Appendix A.3.
Effect on number of parameters: Multihead attention
introduces a 1x1 convolution with Fininput filters and
(2dk+dv) =Fout(2+ )output filters to compute queries,
keys and values and an additional 1x1 convolution with
dv=Fout input and output filters to mix the contribution of different heads. Considering the decrease in filters
in the convolutional part, this leads to the following change
in parameters:
paramsFinFout(2+ (1 k2) +Fout
Fin 2);(5)
where we ignore the parameters introduced by relative position embeddings for simplicity as these are negligible. In
practice, this causes a slight decrease in parameters when
replacing 3x3 convolutions and a slight increase in parameters when replacing 1x1 convolutions. Interestingly, we find | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 16 | practice, this causes a slight decrease in parameters when
replacing 3x3 convolutions and a slight increase in parameters when replacing 1x1 convolutions. Interestingly, we find
in experiments that attention augmented networks still significantly outperform their fully convolutional counterparts
while using less parameters.
Attention Augmented Convolutional Architectures: In
all our experiments, the augmented convolution is followed
by a batch normalization [20] layer which can learn to scale
the contribution of the convolution feature maps and the attention feature maps. We apply our augmented convolution
once per residual block similarly to other visual attention
mechanisms [17, 16, 31, 46] and along the entire architecture as memory permits (see Section 4 for more details).
Since the memory cost O((Nh(HW)2)can be prohibitive for large spatial dimensions, we augment convolutions with attention starting from the last layer (with smallest spatial dimension) until we hit memory constraints. To
reduce the memory footprint of augmented networks, we
typically resort to a smaller batch size and sometimes additionally downsample the inputs to self-attention in the layers with the largest spatial dimensions where it is applied.
Downsampling is performed by applying 3x3 average pooling with stride 2 while the following upsampling (required
for the concatenation) is obtained via bilinear interpolation. | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 17 | Downsampling is performed by applying 3x3 average pooling with stride 2 while the following upsampling (required
for the concatenation) is obtained via bilinear interpolation.
4. Experiments
In the subsequent experiments, we test Attention Augmentation on standard computer vision architectures such
as ResNets [14, 47, 13], and MnasNet [42] on the CIFAR100 [22], ImageNet [9] and COCO [25] datasets. Our experiments show that Attention Augmentation leads to systematic improvements on both image classification and object detection tasks across a broad array of architectures and
computational demands. We validate the utility of the proposed two-dimensional relative attention mechanism in ablation experiments. In all experiments, we substitute convolutional feature maps with self-attention feature maps as
it makes for an easier comparison against the baseline models. Unless specified otherwise, all results correspond to our
two-dimensional relative self-attention mechanism. Experimental details can be found in the Appendix.
4.1. CIFAR-100 image classification
We first investigate how Attention Augmentation performs on CIFAR-100 [22], a standard benchmark for lowresolution imagery, using a Wide ResNet architecture [51].
The Wide-ResNet-28-10 architecture is comprised of 3 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 18 | The Wide-ResNet-28-10 architecture is comprised of 3
stages of 4 residual blocks each using two 33convolutions. We augment the Wide-ResNet-28-10 by augmenting
the first convolution of all residual blocks with relative attention using Nh=8 heads and =2 =0.2 and a minimum of
20 dimensions per head for the keys. We compare Attention
Augmentation (AA) against other forms of attention including Squeeze-and-Excitation (SE) [17] and the parameterfree formulation of Gather-Excite (GE) [16]. Table 1 shows
that Attention Augmentation improves performance both
over the baseline network and Squeeze-and-Excitation at a
similar parameter and complexity cost.
Architecture Params GFlops top-1 top-5
Wide-ResNet [51] 36.3M 10.4 80.3 95.0
GE-Wide-ResNet [16] 36.3M 10.4 79.8 95.0
SE-Wide-ResNet [17] 36.5M 10.4 81.0 95.3
AA-Wide-ResNet (ours) 36.2M 10.9 81.6 95.2
Table 1. Image classification on the CIFAR-100 dataset [22] using | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 19 | Table 1. Image classification on the CIFAR-100 dataset [22] using
the Wide-ResNet 28-10 architecture [51].
4.2. ImageNet image classification with ResNet
We next examine how Attention Augmentation performs
on ImageNet [9, 21], a standard large-scale dataset for high
resolution imagery, across an array of architectures. We
start with the ResNet architecture [14, 47, 13] because of itswidespread use and its ability to easily scale across several
computational budgets. The building block in ResNet-34
comprises two 3x3 convolutions with the same number of
output filters. ResNet-50 and its larger counterparts use a
bottleneck block comprising of 1x1, 3x3, 1x1 convolutions
where the last pointwise convolution expands the number
of filters and the first one contracts the number of filters.
We modify all ResNets by augmenting the 3x3 convolutions as this decreases number of parameters.2We apply
Attention Augmentation in each residual block of the last 3
stages of the architecture – when the spatial dimensions of | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 20 | Attention Augmentation in each residual block of the last 3
stages of the architecture – when the spatial dimensions of
the activation maps are 28x28, 14x14 and 7x7 – and downsample only during the first stage. All attention augmented
networks use =2 =0.2, except for ResNet-34 which uses
= =0.25. The number of attention heads is fixed to Nh=8.
Architecture Params (M) Infer Train top-1
ResNet-50 25.6 - - 76.4
SE [17] 28.1 +12% +92% 77.5 (77.0)
BAM [31] 25.9 +19% +43% 77.3
CBAM [46] 28.1 +56% +132% 77.4 (77.4)
GALA [28] 29.4 +86% +133% 77.5 (77.3)
AA ( = 0:25) 24.3 +29% +25% 77.7
Table 2. Image classification performance of different attention
mechanisms on the ImageNet dataset. refers to the increase
in latency times compared to the ResNet50 on a single Tesla V100 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 21 | mechanisms on the ImageNet dataset. refers to the increase
in latency times compared to the ResNet50 on a single Tesla V100
GPU with Tensorflow using a batch size of 128. For fair comparison, we also include top-1 results (in parentheses) when scaling
networks in width to match 25:6M parameters as the ResNet50
baseline.
Architecture GFlops Params top-1 top-5
ResNet-34 [14] 7.4 21.8M 73.6 91.5
SE-ResNet-34 [17] 7.4 22.0M 74.3 91.8
AA-ResNet-34 (ours) 7.1 20.7M 74.7 92.0
ResNet-50 [14] 8.2 25.6M 76.4 93.1
SE-ResNet-50 [17] 8.2 28.1M 77.5 93.7
AA-ResNet-50 (ours) 8.3 25.8M 77.7 93.8
ResNet-101 [14] 15.6 44.5M 77.9 94.0
SE-ResNet-101 [17] 15.6 49.3M 78.4 94.2 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 22 | SE-ResNet-101 [17] 15.6 49.3M 78.4 94.2
AA-ResNet-101 (ours) 16.1 45.4M 78.7 94.4
ResNet-152 [14] 23.0 60.2M 78.4 94.2
SE-ResNet-152 [17] 23.1 66.8M 78.9 94.5
AA-ResNet-152 (ours) 23.8 61.6M 79.1 94.6
Table 3. Image classification on the ImageNet dataset [9] across
a range of ResNet architectures: ResNet-34, ResNet-50, Resnet101, and ResNet-152 [14, 47, 13].
Table 2 benchmarks Attention Augmentation against
channel and spatial attention mechanisms BAM [31],
CBAM [46] and GALA [28] with channel reduction ratio= 16 on the ResNet50 architecture. Despite the
2We found that augmenting the pointwise expansions works just as well
but does not save parameters or computations.
Architecture GFlops Params top-1 top-5
MnasNet-0.75 0.45 2.91M 73.3 91.3
AA-MnasNet-0.75 0.51 3.02M 73.9 91.6 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 23 | AA-MnasNet-0.75 0.51 3.02M 73.9 91.6
MnasNet-1.0 0.63 3.89M 75.2 92.4
AA-MnasNet-1.0 0.70 4.06M 75.7 92.6
MnasNet-1.25 1.01 5.26M 76.7 93.2
AA-MnasNet-1.25 1.11 5.53M 77.2 93.6
MnasNet-1.4 1.17 6.10M 77.2 93.5
AA-MnasNet-1.4 1.29 6.44M 77.7 93.8
Table 4. Baseline and attention augmented MnasNet [42] accuracies with width multipliers 0.75, 1.0, 1.25 and 1.4.
lack of specialized kernels (See Appendix A.3), Attention
Augmentation offers a competitive accuracy/computational
trade-off compared to previously proposed attention mechanisms. Table 3 compares the non-augmented networks and
Squeeze-and-Excitation (SE) [17] across different network
scales. In all experiments, Attention Augmentation significantly increases performance over the non-augmented
baseline and notably outperforms Squeeze-and-Excitation | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 24 | baseline and notably outperforms Squeeze-and-Excitation
(SE) [17] while being more parameter efficient (Figure 1).
Remarkably, our AA-ResNet-50 performs comparably to
the baseline ResNet-101 and our AA-ResNet-101 outperforms the baseline ResNet-152. These results suggest that
attention augmentation is preferable to simply making networks deeper. We include and discuss attention maps visualizations from different pixel positions in the appendix.
4.3. ImageNet classification with MnasNet
In this section, we inspect the use of Attention Augmentation in a resource constrained setting by conducting
ImageNet experiments with the MnasNet architecture [42],
which is an extremely parameter-efficient architecture. In
particular, the MnasNet was found by neural architecture search [54], using only the highly optimized mobile inverted bottleneck block [36] and the Squeeze-andExcitation operation [17] as the primitives in its search
space. We apply Attention Augmentation to the mobile
inverted bottleneck by replacing convolutional channels in
the expansion pointwise convolution using =2 =0.1 and
Nh=4 heads. Our augmented MnasNets use augmented inverted bottlenecks in the last 13 blocks out of 18 in the | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 25 | Nh=4 heads. Our augmented MnasNets use augmented inverted bottlenecks in the last 13 blocks out of 18 in the
MnasNet architecture, starting when the spatial dimension
is 28x28. We downsample only in the first stage where Attention Augmentation is applied. We leave the final pointwise convolution, also referred to as the “head”, unchanged.
In Table 4, we report ImageNet accuracies for the baseline MnasNet and its attention augmented variants at different width multipliers. Our experiments show that Attention Augmentation yields accuracy improvements across
all width multipliers. Augmenting MnasNets with relative
self-attention incurs a slight parameter increase, however
2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0
number of parameters (millions)737475767778top-1 accuracy
73.3%75.2%76.7%77.2%
73.9%75.7%77.2%77.7%
MnasNet
AA-MnasNet (ours)Figure 3. ImageNet top-1 accuracy as a function of number of parameters for MnasNet (black) and Attention-Augmented-MnasNet | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 26 | AA-MnasNet (ours)Figure 3. ImageNet top-1 accuracy as a function of number of parameters for MnasNet (black) and Attention-Augmented-MnasNet
(red) with width multipliers 0:75,1:0,1:25and1:4.
we verify in Figure 3 that the accuracy improvements are
not just explained by the parameter increase. Additionally,
we note that the MnasNet architecture employs Squeezeand-Excitation at multiple locations that were optimally selected via architecture search, further suggesting the benefits of our method.
4.4. Object Detection with COCO dataset
We next investigate the use of Attention Augmentation
on the task of object detection on the COCO dataset [27].
We employ the RetinaNet architecture with a ResNet-50
and ResNet-101 backbone as done in [26], using the opensourced RetinaNet codebase.3We apply Attention Augmentation uniquely on the ResNet backbone, modifying
them similarly as in our ImageNet classification experiments.
Our relative self-attention mechanism improves the performance of the RetinaNet on both ResNet-50 and ResNet101 as shown in Table 5. Most notably, Attention Augmentation yields a 1.4% mAP improvement over a strong
RetinaNet baseline from [26]. In contrast to the success | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 27 | RetinaNet baseline from [26]. In contrast to the success
of Squeeze-and-Excitation in image classification with ImageNet, our experiments show that adding Squeeze-andExcitation operators in the backbone network of the RetinaNet significantly hurts performance, in spite of grid
searching over the squeeze ratio 2f4;8;16g. We hypothesize that localization requires precise spatial information which SE discards during the spatial pooling operation,
thereby negatively affecting performance. Self-attention on
the other hand maintains spatial information and is likely to
3https://github.com/tensorflow/tpu/tree/master/
models/official/retinanet
Backbone architecture GFlops Params mAP COCO mAP 50 mAP 75
ResNet-50 [26] 182 33.4M 36.8 54.5 39.5
SE-ResNet-50 [17] 183 35.9M 36.5 54.0 39.1
AA-ResNet-50 (ours) 182 33.1M 38.2 56.5 40.7
ResNet-101 [26] 243 52.4M 38.5 56.4 41.2
SE-ResNet-101 [17] 243 57.2M 37.4 55.0 39.9 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 28 | SE-ResNet-101 [17] 243 57.2M 37.4 55.0 39.9
AA-ResNet-101 (ours) 245 51.7M 39.2 57.8 41.9
Table 5. Object detection on the COCO dataset [27] using the RetinaNet architecture [26] with different backbone architectures. We report
mean Average Precision at three different IoU values.
be able to identify object boundaries successfully. Visualizations of attention maps (See Figures 9 and 10 in the Appendix) reveal that some heads are indeed delineating objects from their background which might be important for
localization.
4.5. Ablation Study
Fully-attentional vision models: In this section, we investigate the performance of Attention Augmentation as a
function of the fraction of attentional channels. As we increase this fraction to 100%, we begin to replace a ConvNet with a fully attentional model, only leaving pointwise
convolutions and the stem unchanged. Table 6 presents the
performance of Attention Augmentation on the ResNet-50
architecture for varying ratios = 2f0:25;0:5;0:75;1:0g.
Performance slightly degrades as the ratio of attentional
channels increases, which we hypothesize is partly explained by the average pooling operation for downsampling | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 29 | Performance slightly degrades as the ratio of attentional
channels increases, which we hypothesize is partly explained by the average pooling operation for downsampling
at the first stage where Attention Augmentation is applied.
Attention Augmentation proves however quite robust to the
fraction of attentional channels. For instance, AA-ResNet50 with= =0.75 outperforms its ResNet-50 counterpart,
while being more parameter and flops efficient, indicating
that mostly employing attentional channels is readily competitive.
Perhaps surprisingly, these experiments also reveal that
our proposed self-attention mechanism is a powerful standalone computational primitive for image classification and
that fully attentional models are viable for discriminative visual tasks. In particular, AA-ResNet-50 with = =1, which
uses exclusively attentional channels, is only 2.5% worse
in accuracy than its fully convolutional counterpart, in spite
of downsampling with average pooling and having 25% less
parameters. Notably, this fully attentional architecture4also
outperforms ResNet-34 while being more parameter and
flops efficient (see Table 6).
4We consider pointwise convolutions as dense layers. This architecture | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 30 | flops efficient (see Table 6).
4We consider pointwise convolutions as dense layers. This architecture
employs 4 non-pointwise convolutions in the stem and the first stage of the
architecture, but we believe such operations can be replaced by attention
too.Architecture GFlops Params top-1 top-5
ResNet-34 [14] 7.4 21.8M 73.6 91.5
ResNet-50 [14] 8.2 25.6M 76.4 93.1
= = 0:25 7.9 24.3M 77.7 93.8
= = 0:5 7.3 22.3M 77.3 93.6
= = 0:75 6.8 20.7M 76.7 93.2
= = 1:0 6.3 19.4M 73.9 91.5
Table 6. Attention Augmented ResNet-50 with varying ratios of
attentional channels.
0 25 50 75 100
fraction of attention versus convolutional channels707274767880top-1 accuracy76.4%77.5%
76.6%
75.4%
71.1%77.7%
77.3%
76.7%
73.9% | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 31 | 76.6%
75.4%
71.1%77.7%
77.3%
76.7%
73.9%
no position
with position
Figure 4. Effect of relative position embeddings as the ratio
of attentional channels increases on our Attention-Augmented
ResNet50.
Importance of position encodings: In Figure 4, we show
the effect of our proposed two-dimensional relative position encodings as a function of the fraction of attentional
channels. As expected, experiments demonstrate that our
relative position encodings become increasingly more important as the architecture employs more attentional channels. In particular, the fully self-attentional ResNet-50 gains
2.8% top-1 ImageNet accuracy when using relative position
encodings, which indicates the necessity of maintaining position information for fully self-attentional vision models.
We additionally compare our proposed two-dimensional
Architecture Position Encodings top-1 top-5
AA-ResNet-34 None 74.4 91.9
AA-ResNet-34 2d Sine 74.4 92.0
AA-ResNet-34 CoordConv 74.4 92.0
AA-ResNet-34 Relative (ours) 74.7 92.0
AA-ResNet-50 None 77.5 93.7 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 32 | AA-ResNet-34 Relative (ours) 74.7 92.0
AA-ResNet-50 None 77.5 93.7
AA-ResNet-50 2d Sine 77.5 93.7
AA-ResNet-50 CoordConv 77.5 93.8
AA-ResNet-50 Relative (ours) 77.7 93.8
Table 7. Effects of different position encodings in Attention Augmentation on ImageNet classification.
Position Encodings mAP COCO mAP 50 mAP 75
None 37.7 56.0 40.2
CoordConv [29] 37.4 55.5 40.1
Relative (ours) 38.2 56.5 40.7
Table 8. Effects of different position encodings in Attention Augmentation on the COCO object detection task using a RetinaNet
AA-ResNet-50 backbone.
relative position encodings to other position encoding
schemes. We apply Attention Augmentation using the same
hyperparameters as 4.2 with the following different position encoding schemes: 1) The position-unaware version of
self-attention (referred to as None ), 2) a two-dimensional
implementation of the sinusoidal positional waves (referred | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 33 | self-attention (referred to as None ), 2) a two-dimensional
implementation of the sinusoidal positional waves (referred
to as 2d Sine ) as used in [32], 3) CoordConv [29] for which
we concatenate (x,y,r) coordinate channels to the inputs of
the attention function, and 4) our proposed two-dimensional
relative position encodings (referred to as Relative ).
In Table 7 and 8, we present the results on ImageNet
classification and the COCO object detection task respectively. On both tasks, Attention Augmentation without position encodings already yields improvements over the fully
convolutional non-augmented variants. Our experiments
also reveal that the sinusoidal encodings and the coordinate
convolution do not provide improvements over the positionunaware version of Attention Augmentation. We obtain additional improvements when using our two-dimensional relative attention, demonstrating the utility of preserving translation equivariance while preventing permutation equivariance.
5. Discussion and future work
In this work, we consider the use of self-attention for vision models as an alternative to convolutions. We introduce
a novel two-dimensional relative self-attention mechanism
for images that enables training of competitive fully selfattentional vision models on image classification for the first | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 34 | for images that enables training of competitive fully selfattentional vision models on image classification for the first
time. We propose to augment convolutional operators with
this self-attention mechanism and validate the superiority ofthis approach over other attention schemes. Extensive experiments show that Attention Augmentation leads to systematic improvements on both image classification and object detection tasks across a wide range of architectures and
computational settings.
Several open questions from this work remain. In future work, we will focus on the fully attentional regime
and explore how different attention mechanisms trade off
computational efficiency versus representational power. For
instance, identifying a local attention mechanism may result in an efficient and scalable computational mechanism
that could prevent the need for downsampling with average
pooling [34]. Additionally, it is plausible that architectural
design choices that are well suited when exclusively relying
on convolutions are suboptimal when using self-attention
mechanisms. As such, it would be interesting to see if using Attention Augmentation as a primitive in automated architecture search procedures proves useful to find even better models than those previously found in image classification [55], object detection [12], image segmentation [6] and
other domains [5, 1, 35, 8]. Finally, one can ask to which | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 35 | other domains [5, 1, 35, 8]. Finally, one can ask to which
degree fully attentional models can replace convolutional
networks for visual tasks.
Acknowledgements
The authors would like to thank Tsung-Yi Lin, Prajit Ramachandran, Mingxing Tan, Yanping Huang and the
Google Brain team for insightful comments and discussions.
References
[1] Maximilian Alber, Irwan Bello, Barret Zoph, Pieter-Jan Kindermans, Prajit Ramachandran, and Quoc V . Le. Backprop
evolution. CoRR , abs/1808.02822, 2018. 8
[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio.
Neural machine translation by jointly learning to align and
translate. In International Conference on Learning Representations , 2015. 2
[3] Irwan Bello, Sayali Kulkarni, Sagar Jain, Craig Boutilier,
Ed Huai-hsin Chi, Elad Eban, Xiyang Luo, Alan Mackey,
and Ofer Meshi. Seq2slate: Re-ranking and slate optimization with rnns. CoRR , abs/1810.02019, 2018. 2 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 36 | and Ofer Meshi. Seq2slate: Re-ranking and slate optimization with rnns. CoRR , abs/1810.02019, 2018. 2
[4] Irwan Bello, Hieu Pham, Quoc V . Le, Mohammad Norouzi,
and Samy Bengio. Neural combinatorial optimization with
reinforcement learning. 2016. 2
[5] Irwan Bello, Barret Zoph, Vijay Vasudevan, and Quoc V . Le.
Neural optimizer search with reinforcement learning. In Proceedings of the 34th International Conference on Machine
Learning - Volume 70 , ICML’17, pages 459–468. JMLR.org,
2017. 8
[6] Liang-Chieh Chen, Maxwell Collins, Yukun Zhu, George
Papandreou, Barret Zoph, Florian Schroff, Hartwig Adam,
and Jon Shlens. Searching for efficient multi-scale architectures for dense image prediction. In Advances in Neural
Information Processing Systems , pages 8713–8724, 2018. 2,
8
[7] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng
Yan, and Jiashi Feng. A2-nets: Double attention networks. | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 37 | [7] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng
Yan, and Jiashi Feng. A2-nets: Double attention networks.
CoRR , abs/1810.11579, 2018. 3
[8] Ekin Dogus Cubuk, Barret Zoph, Dandelion Man ´e, Vijay Vasudevan, and Quoc V . Le. Autoaugment: Learning augmentation policies from data. CoRR , abs/1805.09501, 2018. 8
[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,
and Li Fei-Fei. Imagenet: A large-scale hierarchical image
database. In IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 2009. 1, 2, 5
[10] Xavier Gastaldi. Shake-shake regularization. arXiv preprint
arXiv:1705.07485 , 2017. 11
[11] Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. Dropblock:
A regularization method for convolutional networks. In
Advances in Neural Information Processing Systems , pages
10750–10760, 2018. 11 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 38 | A regularization method for convolutional networks. In
Advances in Neural Information Processing Systems , pages
10750–10760, 2018. 11
[12] Golnaz Ghiasi, Tsung-Yi Lin, Ruoming Pang, and Quoc V
Le. NAS-FPN: Learning scalable feature pyramid architecture for object detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2019. 8
[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition , 2016.
2, 5
[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Identity mappings in deep residual networks. In European
Conference on Computer Vision , 2016. 1, 2, 5, 7, 11
[15] Sepp Hochreiter and Juergen Schmidhuber. Long short-term
memory. Neural Computation , 1997. 2
[16] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Andrea
Vedaldi. Gather-excite: Exploiting feature context in convolutional neural networks. In Advances in Neural Information
Processing Systems , pages 9423–9433, 2018. 3, 4, 5 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 39 | Processing Systems , pages 9423–9433, 2018. 3, 4, 5
[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition , 2018. 1, 2, 3, 4, 5, 6, 7
[18] Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Curtis Hawthorne, Andrew M Dai,
Matthew D Hoffman, and Douglas Eck. Music transformer.
InAdvances in Neural Processing Systems , 2018. 3, 4
[19] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu,
Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In
IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2
[20] Sergey Ioffe and Christian Szegedy. Batch normalization:
Accelerating deep network training by reducing internal covariate shift. In International Conference on Learning Representations , 2015. 4 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 40 | Accelerating deep network training by reducing internal covariate shift. In International Conference on Learning Representations , 2015. 4
[21] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2, 5[22] Alex Krizhevsky. Learning multiple layers of features from
tiny images. Technical report, University of Toronto, 2009.
2, 5
[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton.
Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing System, 2012. 1, 2
[24] Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick
Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE , 1998. 1
[25] Tsung-Yi Lin, Piotr Doll ´ar, Ross Girshick, Kaiming He,
Bharath Hariharan, and Serge Belongie. Feature pyramid
networks for object detection. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition ,
2017. 5 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 41 | networks for object detection. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition ,
2017. 5
[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and
Piotr Doll ´ar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer
vision , pages 2980–2988, 2017. 6, 7, 11
[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays,
Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C Lawrence
Zitnick. Microsoft coco: Common objects in context. In
European Conference on Computer Vision , pages 740–755.
Springer, 2014. 2, 6, 7
[28] Drew Linsley, Dan Scheibler, Sven Eberhardt, and Thomas
Serre. Global-and-local attention networks for visual recognition. CoRR , abs/1805.08819, 2018. 5
[29] Rosanne Liu, Joel Lehman, Piero Molino, Felipe Petroski
Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An
intriguing failing of convolutional neural networks and the | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 42 | Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An
intriguing failing of convolutional neural networks and the
coordconv solution. In Advances in Neural Information Processing Systems , pages 9628–9639, 2018. 3, 8
[30] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint
arXiv:1608.03983 , 2016. 11
[31] Jongchan Park, Sanghyun Woo, Joon-Young Lee, and In So
Kweon. Bam: bottleneck attention module. In British Machine Vision Conference , 2018. 3, 4, 5
[32] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Łukasz
Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine
Learning , 2018. 3, 8
[33] Andrew Rabinovich, Andrea Vedaldi, Carolina Galleguillos,
Eric Wiewiora, and Serge Belongie. Objects in context.
2007. 1
[34] Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan
Bello, Anselm Levskaya, and Jonathon Shlens. Stand-alone | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 43 | Bello, Anselm Levskaya, and Jonathon Shlens. Stand-alone
self-attention in vision models. CoRR , abs/1906.05909,
2019. 8
[35] Prajit Ramachandran, Barret Zoph, and Quoc V . Le. Searching for activation functions. CoRR , abs/1710.05941, 2017.
8
[36] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted
residuals and linear bottlenecks. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition ,
pages 4510–4520, 2018. 6
[37] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Selfattention with relative position representations. arXiv
preprint arXiv:1803.02155 , 2018. 2, 3, 4
[38] David R. So, Chen Liang, and Quoc V . Le. The evolved
transformer. CoRR , abs/1901.11117, 2019. 3
[39] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and
Alex Alemi. Inception-v4, Inception-Resnet and the impact | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 44 | [39] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and
Alex Alemi. Inception-v4, Inception-Resnet and the impact
of residual connections on learning. In International Conference on Learning Representations Workshop Track , 2016.
2
[40] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet,
Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent
Vanhoucke, and Andrew Rabinovich. Going deeper with
convolutions. In IEEE Conference on Computer Vision and
Pattern Recognition , 2015. 2
[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon
Shlens, and Zbigniew Wojna. Rethinking the Inception architecture for computer vision. In IEEE Conference on Computer Vision and Pattern Recognition , 2016. 2, 11
[42] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan,
and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018. 2,
5, 6, 11 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 45 | 5, 6, 11
[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia
Polosukhin. Attention is all you need. In Advances in Neural
Information Processing Systems , pages 5998–6008, 2017. 1,
2, 3
[44] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer
networks. In NIPS , pages 2692–2700, 2015. 2
[45] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 7794–7803, 2018. 3
[46] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In
So Kweon. Cbam: Convolutional block attention module.
InProceedings of the European Conference on Computer Vision (ECCV) , pages 3–19, 2018. 3, 4, 5
[47] Saining Xie, Ross Girshick, Piotr Doll ´ar, Zhuowen Tu, and
Kaiming He. Aggregated residual transformations for deep | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 46 | [47] Saining Xie, Ross Girshick, Piotr Doll ´ar, Zhuowen Tu, and
Kaiming He. Aggregated residual transformations for deep
neural networks. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition , 2017. 2, 5
[48] Yoshihiro Yamada, Masakazu Iwamura, Takuya Akiba, and
Koichi Kise. Shakedrop regularization for deep residual
learning. arXiv preprint arXiv:1802.02375 , 2018. 11
[49] Baosong Yang, Longyue Wang, Derek F. Wong, Lidia S.
Chao, and Zhaopeng Tu. Convolutional self-attention network. In CoRR , volume abs/1810.13320, 2018. 2
[50] Adams Wei Yu, David Dohan, Minh-Thang Luong, Rui
Zhao, Kai Chen, Mohammad Norouzi, and Quoc V Le.
QAnet: Combining local convolution with global selfattention for reading comprehension. In International Conference on Learning Representations , 2018. 2 | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 47 | QAnet: Combining local convolution with global selfattention for reading comprehension. In International Conference on Learning Representations , 2018. 2
[51] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference , 2016. 5[52] Vinicius Zambaldi, David Raposo, Adam Santoro, Victor Bapst, Yujia Li, Igor Babuschkin, Karl Tuyls, David
Reichert, Timothy Lillicrap, Edward Lockhart, Murray
Shanahan, Victoria Langston, Razvan Pascanu, Matthew
Botvinick, Oriol Vinyals, and Peter Battaglia. Deep reinforcement learning with relational inductive biases. In ICLR ,
2019. 2
[53] Han Zhang, Ian J. Goodfellow, Dimitris N. Metaxas, and
Augustus Odena. Self-attention generative adversarial networks. arXiv:1805.08318 , 2018. 3
[54] Barret Zoph and Quoc V . Le. Neural architecture search
with reinforcement learning. In International Conference on
Learning Representations , 2017. 6
[55] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V
Le. Learning transferable architectures for scalable image | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 48 | [55] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V
Le. Learning transferable architectures for scalable image
recognition. In Proceedings of the IEEE conference on
computer vision and pattern recognition , pages 8697–8710,
2018. 2, 8, 11
A. Appendix
A.1. Experimental details
Tuning Unless specified otherwise, we use the default hyperparameters found in reference baseline codebases without tuning.was searched inf0.1, 0.2, 0.5g, inf0.0, 0.1,
0.25, 0.5, 0.75, 1.0 gand the number of heads was chosen
based on memory constraints (starting from 8 and decreasing when necessary). We report the final accuracy for each
run without performing early stopping.
CIFAR-100 Given the low resolution of CIFAR-100 images, we do not downsample feature maps before the attention operation and instead resort to a smaller batch size.
We train all networks for 500 epochs using synchronous
SGD with momentum 0.9 distributed across 8 TESLA V100
GPUs. The learning rate is linearly scaled from 0 to
0:2B=256, whereBis the total batch size, for the first 5% | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 49 | GPUs. The learning rate is linearly scaled from 0 to
0:2B=256, whereBis the total batch size, for the first 5%
training epochs and then annealed with cosine decay [30].
We use standard CIFAR preprocessing: mean normalizing,
random flipping and cropping [55, 10, 48]. Non-augmented
architectures are trained with a batch size of 1024 and a
weight decay of 2e-4. Augmented architectures are trained
with batch size of 256 and a weight decay of 5e-4.
ImageNet classification with ResNet We train all
ResNet architectures for 100 epochs using synchronous
SGD with momentum 0.9 across 8 TESLA V100 GPUs and
weight decay of 1e-4. We use the largest batch size per
workerB2f32;64;128;256gthat fits in a minibatch. The
initial learning rate is scaled linearly according to the total
batch size using a base learning rate of 0.128 for total batch
size of 256. During training, we linearly scale the learning
rate from 0 to this value for the first 5% of training epochs
and divide it by 10 at epochs 30, 60, 80 and 90. We use | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 50 | rate from 0 to this value for the first 5% of training epochs
and divide it by 10 at epochs 30, 60, 80 and 90. We use
standard Inception data augmentation as described in [41].
ImageNet classification with MnasNet We follow the
training setup described in [42] and train all networks for
350 epochs with the RMSProp optimizer using exponential
learning rate decay. When training our augmented MnasNets, we divide the learning rate by 2 and adjusted the learning rate decay so that the final learning rate stays the same.
Object Detection with COCO dataset We follow the
setup described in [26, 11] and train the RetinaNet from
scratch for 150 epochs without using ImageNet pretraining
for the ResNet backbone.
A.2. Computational & Memory costs
Table 9 provides the breakdown of self-attention related computational costs per image. Storing attention
maps in each layer induces a memory cost of Nh(HW)2bfloat16 . At inference, the memory cost for storing attention maps is only 1.2% of the memory required to store
model parameters (49MB).
Layer Memory Params FLOPS
fStage 2 - H=W=14g* 4 600KB 43k 22M | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 51 | model parameters (49MB).
Layer Memory Params FLOPS
fStage 2 - H=W=14g* 4 600KB 43k 22M
fStage 3 - H=W=14g* 6 600KB 90k 40M
fStage 4 - H=W=7g* 3 37.5KB 190k 19M
Training 6MB (total) 1.3M 390M
Inference 600KB (max) 1.3M 390M
Table 9. Computational costs associated with self-attention in the
forward pass of the ResNet50. During inference, we only consider
the largest memory cost since activations are not stored.
Figures 5 and 6 show the accuracies of our attention
augmented networks across FLOPS counts, which correlate
with running times across hardware platforms.
5 10 15 20 25
computational demand (GFlops).72737475767778798081top-1 accuracy.
73.6%76.4%77.9%78.4%
74.3%77.5%78.4%78.9%
74.7%77.7%78.7%79.1%
ResNet
SE-ResNet
AA-ResNet (ours)
Figure 5. ImageNet top-1 accuracy as a function of computational | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 52 | ResNet
SE-ResNet
AA-ResNet (ours)
Figure 5. ImageNet top-1 accuracy as a function of computational
demand for variety of ResNet architectures [14]. From left to right:
ResNet-34, ResNet-50, ResNet-101 and ResNet-152.
0.4 0.6 0.8 1.0 1.2 1.4
computational demand (GFlops).737475767778top-1 accuracy
73.3%75.2%76.7%77.2%
73.9%75.7%77.2%77.7%
MnasNet
AA-MnasNet (ours)
Figure 6. ImageNet top-1 accuracy as a function of computational
demand for MnasNet (black) and Attention-Augmented-MnasNet
(red) with width multipliers 0:75,1:0,1:25and1:4.
A.3. 2D Relative Self-Attention implementation
While our method is simple and only requires matrix
multiplication, addition and the softmax operation (Equations 3 and 4), our implementation relies on non-trivial operations (e.g. tiling, transposing and reshaping) because
no low-level kernels currently exist for hardware platforms.
Future work may develop specialized kernels as previously | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 53 | no low-level kernels currently exist for hardware platforms.
Future work may develop specialized kernels as previously
done for convolutions. Therefore, we believe that current
latency times (Table 2) reflect the lack of dedicated engineering as opposed to inefficiency in the proposed method.
def shape list(x):
"""Return list of dims, statically where possible."""
static = x.get shape().as list()
shape = tf.shape(x)
ret = []
for i, static dim in enumerate(static):
dim = static dim or shape[i]
ret.append(dim)
return ret
def split heads 2d(inputs, Nh):
"""Split channels into multiple heads."""
B, H, W, d = shape list(inputs)
retshape = [B, H, W, Nh, d // Nh]
split = tf.reshape(inputs, ret shape)
return tf.transpose(split, [0, 3, 1, 2, 4])
def combine heads 2d(inputs):
"""Combine heads (inverse of split heads 2d)."""
transposed = tf.transpose(inputs, [0, 2, 3, 1, 4])
Nh, channels = shape list(transposed)[ 2:] | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 54 | transposed = tf.transpose(inputs, [0, 2, 3, 1, 4])
Nh, channels = shape list(transposed)[ 2:]
retshape = shape list(transposed)[: 2] + [Nh channels]
return tf.reshape(transposed, ret shape)
def rel toabs(x):
"""Converts tensor from relative to aboslute indexing."""
# [B, Nh, L, 2L 1]
B, Nh, L, = shape list(x)
# Pad to shift from relative to absolute indexing.
colpad = tf.zeros((B, Nh, L, 1))
x = tf.concat([x, col pad], axis=3)
flat x = tf.reshape(x, [B, Nh, L 2L])
flat pad = tf.zeros((B, Nh, L 1))
flat xpadded = tf.concat([flat x, flat pad], axis=2)
# Reshape and slice out the padded elements.
final x = tf.reshape(flat xpadded, [B, Nh, L+1, 2 L 1]) | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 55 | final x = tf.reshape(flat xpadded, [B, Nh, L+1, 2 L 1])
final x = final x[:, :, :L, L 1:]
return final x
def relative logits 1d(q, rel k, H, W, Nh, transpose mask):
"""Compute relative logits along one dimenion."""
rellogits = tf.einsum(’bhxyd,md >bhxym’, q, rel k)
# Collapse height and heads
rellogits = tf.reshape(
rellogits, [ 1, Nh H, W, 2 W 1])
rellogits = rel toabs(rel logits)
# Shape it and tile height times
rellogits = tf.reshape(rel logits, [ 1, Nh, H, W, W])
rellogits = tf.expand dims(rel logits, axis=3)
rellogits = tf.tile(rel logits, [1, 1, 1, H, 1, 1])
# Reshape for adding to the logits.
rellogits = tf.transpose(rel logits, transpose mask) | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 56 | # Reshape for adding to the logits.
rellogits = tf.transpose(rel logits, transpose mask)
rellogits = tf.reshape(rel logits, [ 1, Nh, H W, HW])
return rel logits
Figure 7. Helper functions in Tensorflow for 2D relative selfattention.def relative logits(q, H, W, Nh, dkh):
"""Compute relative logits."""
# Relative logits in width dimension first.
relembeddings w = tf.get variable(
’rwidth’, shape=(2 W 1, dkh),
initializer=tf.random normal initializer(dkh 0.5))
# [B, Nh, HW, HW]
rellogits w = relative logits 1d(
q, rel embeddings w, H, W, Nh, [0, 1, 2, 4, 3, 5])
# Relative logits in height dimension next.
# For ease, we 1) transpose height and width,
# 2) repeat the above steps and
# 3) transpose to eventually put the logits
# in their right positions.
relembeddings h = tf.get variable( | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 57 | # 2) repeat the above steps and
# 3) transpose to eventually put the logits
# in their right positions.
relembeddings h = tf.get variable(
’rheight’, shape=(2 H 1, dkh),
initializer=tf.random normal initializer(dkh 0.5))
# [B, Nh, HW, HW]
rellogits h = relative logits 1d(
tf.transpose(q, [0, 1, 3, 2, 4]),
relembeddings h, W, H, Nh, [0, 1, 4, 2, 5, 3])
return rel logits h, rel logits w
def self attention 2d(inputs, dk, dv, Nh, relative=True):
"""2d relative self attention."""
, H, W, = shape list(inputs)
dkh = dk // Nh
dvh = dv // Nh
flatten hw = lambda x, d: tf.reshape(x, [ 1, Nh, H W, d])
# Compute q, k, v
kqv = tf.layers.conv2d(inputs, 2 dk + dv, 1) | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 58 | # Compute q, k, v
kqv = tf.layers.conv2d(inputs, 2 dk + dv, 1)
k, q, v = tf.split(kqv, [dk, dk, dv], axis=3)
q= dkh 0.5 # scaled dot product
# After splitting, shape is [B, Nh, H, W, dkh or dvh]
q = split heads 2d(q, Nh)
k = split heads 2d(k, Nh)
v = split heads 2d(v, Nh)
# [B, Nh, HW, HW]
logits = tf.matmul(flatten hw(q, dkh), flatten hw(k, dkh),
transpose b=True)
if relative:
rellogits h, rel logits w = relative logits(q, H, W, Nh,
dkh)
logits += rel logits h
logits += rel logits w
weights = tf.nn.softmax(logits)
attn out = tf.matmul(weights, flatten hw(v, dvh)) | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 59 | weights = tf.nn.softmax(logits)
attn out = tf.matmul(weights, flatten hw(v, dvh))
attn out = tf.reshape(attn out, [ 1, Nh, H, W, dvh])
attn out = combine heads 2d(attn out)
# Project heads
attn out = tf.layers.conv2d(attn out, dv, 1)
return attn out
def augmented conv2d(X, Fout, k, dk, dv, Nh, relative):
conv out = tf.layers.conv2d(inputs=X, filters=Fout dv,
kernel size=k, padding=’same’)
attn out = self attention 2d(X, dk, dv, Nh, relative=
relative)
return tf.concat([conv out, attn out], axis=3)
Figure 8. Tensorflow code for 2D relative self-attention.
A.4. Attention visualizations.
In Figure 10, we present attention maps visualizations
for the input image shown in Figure 9. We see that attention
heads learn to specialize to different content and notably can
delineate object boundaries. | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
1904.09925 | 60 | for the input image shown in Figure 9. We see that attention
heads learn to specialize to different content and notably can
delineate object boundaries.
Figure 9. An input image. The red crosses indexed 1 to 4 represent
the pixel locations for which we show the attention maps in Figure
10.
Figure 10. Visualization of attention maps for an augmented convolution in the Attention-Augmented-ResNet50. Rows correspond to the 8 different heads and columns correspond to the 4
pixel locations depicted in the input image (See Figure 9). | 1904.09925 | Attention Augmented Convolutional Networks | Convolutional networks have been the paradigm of choice in many computer
vision applications. The convolution operation however has a significant
weakness in that it only operates on a local neighborhood, thus missing global
information. Self-attention, on the other hand, has emerged as a recent advance
to capture long range interactions, but has mostly been applied to sequence
modeling and generative modeling tasks. In this paper, we consider the use of
self-attention for discriminative visual tasks as an alternative to
convolutions. We introduce a novel two-dimensional relative self-attention
mechanism that proves competitive in replacing convolutions as a stand-alone
computational primitive for image classification. We find in control
experiments that the best results are obtained when combining both convolutions
and self-attention. We therefore propose to augment convolutional operators
with this self-attention mechanism by concatenating convolutional feature maps
with a set of feature maps produced via self-attention. Extensive experiments
show that Attention Augmentation leads to consistent improvements in image
classification on ImageNet and object detection on COCO across many different
models and scales, including ResNets and a state-of-the art mobile constrained
network, while keeping the number of parameters similar. In particular, our
method achieves a $1.3\%$ top-1 accuracy improvement on ImageNet classification
over a ResNet50 baseline and outperforms other attention mechanisms for images
such as Squeeze-and-Excitation. It also achieves an improvement of 1.4 mAP in
COCO Object Detection on top of a RetinaNet baseline. | http://arxiv.org/pdf/1904.09925 | [
"Irwan Bello",
"Barret Zoph",
"Ashish Vaswani",
"Jonathon Shlens",
"Quoc V. Le"
] | [
"cs.CV"
] | ICCV 2019 | null | cs.CV | 20190422 | 20200909 | [] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.