doi
stringlengths 10
10
| chunk-id
int64 0
936
| chunk
stringlengths 401
2.02k
| id
stringlengths 12
14
| title
stringlengths 8
162
| summary
stringlengths 228
1.92k
| source
stringlengths 31
31
| authors
stringlengths 7
6.97k
| categories
stringlengths 5
107
| comment
stringlengths 4
398
⌀ | journal_ref
stringlengths 8
194
⌀ | primary_category
stringlengths 5
17
| published
stringlengths 8
8
| updated
stringlengths 8
8
| references
list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1703.10135 | 7 | Char2Wav (Sotelo et al., 2017) is an independently-developed end-to-end model that can be trained on characters. However, Char2Wav still predicts vocoder parameters before using a SampleRNN neural vocoder (Mehri et al., 2016), whereas Tacotron directly predicts raw spectrogram. Also, their seq2seq and SampleRNN models need to be separately pre-trained, but our model can be trained
1Sound demos can be found at https://google.github.io/tacotron
2
# target
from scratch. Finally, we made several key modiï¬cations to the vanilla seq2seq paradigm. As shown later, a vanilla seq2seq model does not work well for character-level inputs.
# 3 MODEL ARCHITECTURE
The backbone of Tacotron is a seq2seq model with attention (Bahdanau et al., 2014; Vinyals et al., 2015). Figure 1 depicts the model, which includes an encoder, an attention-based decoder, and a post-processing net. At a high-level, our model takes characters as input and produces spectrogram frames, which are then converted to waveforms. We describe these components below.
HHL 4 Bidirectional RNN Highway layers Residual connection Conv1D layers 4 Conv1D projections 4 Max-pool along time (stride=1) | 1703.10135#7 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 8 | We propose a novel network architecture that addresses both of these problems through careful design changes, allowing for resource-efï¬cient image classiï¬cation. Our network uses a cascade of intermediate classiï¬ers throughout the network. The ï¬rst problem, of classiï¬ers altering the internal representation, is addressed through the introduction of dense connectivity (Huang et al., 2017). By connecting all layers to all classiï¬ers, features are no longer dominated by the most imminent early- exit and the trade-off between early or later classiï¬cation can be performed elegantly as part of the loss function. The second problem, the lack of coarse-scale features in early layers, is addressed by adopting a multi-scale network structure. At each layer we produce features of all scales (ï¬ne-to- coarse), which facilitates good classiï¬cation early on but also extracts low-level features that only become useful after several more layers of processing. Our network architecture is illustrated in Figure 2, and we refer to it as Multi-Scale DenseNet (MSDNet). | 1703.09844#8 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 8 | network is fully symmetric and embedded in the original network, it lacks the ability of handle heterogeneous agent types. Also it is a single network for all agents, and therefore its scalability is unclear. In this paper, we solve these issues by creating a dedicated bi-directional communi- cation channel using recurrent neural networks (Schuster and Paliwal 1997). As such, heterogeneous agents can be created with a different set of parameters and output actions. The bi-directional nature means that the communication is not entirely symmetric, and the different priority among agents would help solving any possible tie between multiple opti- mal joint actions (Busoniu, Babuska, and De Schutter 2008; Spaan et al. 2002). Multiagent systems have been explored on specific Star- Craft games. Google DeepMind released a game interface based on StarCraft II and claimed that it is hard to make significant progress on the full game even with the state-of- the-art RL algorithms (Vinyals et al. 2017). Usunier et al. presented a heuristic exploration technique for learning deter- ministic policies in micro-management tasks. Both Synnaeve et al. and Usunier et al. focused on a greedy MDP approach, 1.e., the action of an | 1703.10069#8 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 9 | We evaluate MSDNets on three image-classiï¬cation datasets. In the anytime classiï¬cation setting, we show that it is possible to provide the ability to output a prediction at any time while maintain high accuracies throughout. In the budget batch classiï¬cation setting we show that MSDNets can be effectively used to adapt the amount of computation to the difï¬culty of the example to be classiï¬ed, which allows us to reduce the computational requirements of our models drastically whilst perform- ing on par with state-of-the-art CNNs in terms of overall classiï¬cation accuracy. To our knowledge this is the ï¬rst deep learning architecture of its kind that allows dynamic resource adaptation with a single model and obtains competitive results throughout.
2Source: https://opensignal.com/reports/2015/08/android-fragmentation/ 3https://en.wikipedia.org/wiki/Google_Images
2
Published as a conference paper at ICLR 2018
s fam) wee xt £0) AC) features classifier regular conv - 3 eae ONY ea? one h(-) + see layer concatenation strided conv identity | 1703.09844#9 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 9 | in micro-management tasks. Both Synnaeve et al. and Usunier et al. focused on a greedy MDP approach, 1.e., the action of an agent is dependent explicitly on the ac- tion of other agents. In our paper, the dependency of agents is rather modelled over hidden layers by making use of bi- Attention Neron, WB Bicirectionar NN WG Poticy Action (a) Multiagent policy networks (b) Multiagent Q networks Figure 1: Bidirectionally-Coordinated Net (BiCNet). As a means of communication, bi-direction recurrent networks have been used to connect each individual agentâs policy and and Q networks. The learning is done by multiagent deterministic actor-critic as derived in the text. directional RNN (Schuster and Paliwal 1997). A significant benefit over the greedy solution is that, while keeping simple, the communication happens in the latent space so that high- level information can be passed between agents; meanwhile, the gradient updates from all the agents can be efficiently propagated through the entire network. Recently, Foerster et al. has attempted to solve the non- stationarity problem in mutliagent learning by improving the replay buffer, and tested the DIAL model in a way | 1703.10069#9 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 9 | We ï¬rst describe a building block dubbed CBHG, illustrated in Figure 2. CBHG consists of a bank of 1-D convolutional ï¬lters, followed by highway networks (Srivastava et al., 2015) and a bidirectional gated recurrent unit (GRU) (Chung et al., 2014) recurrent neural net (RNN). CBHG is a powerful module for extracting representations from sequences. The input sequence is ï¬rst convolved with K sets of 1-D convolutional ï¬lters, where the k-th set contains Ck ï¬lters of width k (i.e. k = 1, 2, . . . , K). These ï¬lters explicitly model local and contextual information (akin to modeling unigrams, bigrams, up to K-grams). The convolution outputs are stacked together and further max pooled along time to increase local invariances. Note that we use a stride of 1 to preserve the original time resolution. We further pass the processed sequence to a few ï¬xed-width 1-D convolutions, whose outputs are added with the original input sequence via residual connections (He et al., 2016). Batch | 1703.10135#9 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 10 | s fam) wee xt £0) AC) features classifier regular conv - 3 eae ONY ea? one h(-) + see layer concatenation strided conv identity
Figure 2: Illustration of the ï¬rst four layers of an MSDNet with three scales. The horizontal direction cor- responds to the layer direction (depth) of the network. The vertical direction corresponds to the scale of the feature maps. Horizontal arrows indicate a regular convolution operation, whereas diagonal and vertical arrows indicate a strided convolution operation. Classiï¬ers only operate on feature maps at the coarsest scale. Connec- tions across more than one layer are not drawn explicitly: they are implicit through recursive concatenations.
# 2 RELATED WORK
We brieï¬y review related prior work on computation-efï¬cient networks, memory-efï¬cient networks, and resource-sensitive machine learning, from which our network architecture draws inspiration. | 1703.09844#10 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 10 | et al. has attempted to solve the non- stationarity problem in mutliagent learning by improving the replay buffer, and tested the DIAL model in a way that all agents are fully decentralised. The COMA model (Foerster et al. 2017a) was then proposed to tackle the challenge of multiagent credit assignment. Through the introduction of the counterfactual reward; the idea of training multiagent systems in the centralised critic and decentralised actors way was further reinforced. At the same time, the framework of centralised learning and decentralised execution was also adopted by MADDPG in (Lowe et al. 2017) in some simpler, non-startcraft cases. By contrast, our BiCNet makes use of memory to form a communication channel among agents where the parameter space of communication is independent of the number of agents. Multiagent Bidirectionally-Coordinated Nets StartCraft Combat as Stochastic Games The StarCraft combat games, a.k.a., the micromanagement tasks, refer to the low-level, short-term control of the army members during a combat against the enemy members. For each combat, the agents in one side are fully coop- erative, and they compete with the opponents; therefore, | 1703.10069#10 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 10 | few ï¬xed-width 1-D convolutions, whose outputs are added with the original input sequence via residual connections (He et al., 2016). Batch normalization (Ioffe & Szegedy, 2015) is used for all convolutional layers. The convolution outputs are fed into a multi-layer highway network to extract high-level features. Finally, we stack a bidirectional GRU RNN on top to extract sequential features from both forward and backward context. CBHG is inspired from work in machine translation (Lee et al., 2016), where the main differences from Lee et al. (2016) include using non-causal convolutions, batch normalization, residual connections, and stride=1 max pooling. We found that these modiï¬cations improved generalization. | 1703.10135#10 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 11 | Computation-efï¬cient networks. Most prior work on (convolutional) networks that are computa- tionally efï¬cient at test time focuses on reducing model size after training. In particular, many stud- ies propose to prune weights (LeCun et al., 1989; Hassibi et al., 1993; Li et al., 2017) or quantize weights (Hubara et al., 2016; Rastegari et al., 2016) during or after training. These approaches are generally effective because deep networks often have a substantial number of redundant weights that can be pruned or quantized without sacriï¬cing (and sometimes even improving) performance. Prior work also studies approaches that directly learn compact models with less parameter redundancy. For example, the knowledge-distillation method (Bucilua et al., 2006; Hinton et al., 2014) trains small student networks to reproduce the output of a much larger teacher network or ensemble. Our work differs from those approaches in that we train a single model that trades off computation for accuracy at test time without any re-training or ï¬netuning. Indeed, weight pruning and knowledge distillation can be used in combination with our approach, and may lead to further improvements. | 1703.09844#11 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 11 | each combat can be considered as a zero-sum competi- tive game between N agents and M enemies. We formu- late it as a zero-sum Stochastic Game (SG) (Owen 1995), i.e., a dynamic game in a multiple state situation played by multiple agents. A SG can be described by a tuple (S, {A}, (BG, T, {RM}. Let S denotes the state space of the current game, shared among all the agents. Initial state s; follows s; ~ p(s). We define the action space of the controlled agent 7 as A;, and the action space of the enemy j as Bj. T : S x AN x BM â S stands for the deterministic transition function of the environment, and Ri: Sx AN x BM â R represents the reward function of each agent i for i ⬠[1, N]. | 1703.10069#11 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 12 | Resource-efï¬cient machine learning. Various prior studies explore computationally efï¬cient vari- ants of traditional machine-learning models (Viola & Jones, 2001; Grubb & Bagnell, 2012; Karayev et al., 2014; Trapeznikov & Saligrama, 2013; Xu et al., 2012; 2013; Nan et al., 2015; Wang et al., 2015). Most of these studies focus on how to incorporate the computational requirements of com- puting particular features in the training of machine-learning models such as (gradient-boosted) decision trees. Whilst our study is certainly inspired by these results, the architecture we explore differs substantially: most prior work exploits characteristics of machine-learning models (such as decision trees) that do not apply to deep networks. Our work is possibly most closely related to recent work on FractalNets (Larsson et al., 2017), which can perform anytime prediction by pro- gressively evaluating subnetworks of the full network. FractalNets differ from our work in that they are not explicitly optimized for computation efï¬ciency and consequently our experiments show that MSDNets substantially outperform FractalNets. Our | 1703.09844#12 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 12 | In order to maintain a flexible framework that could allow an arbitrary number of agents, we consider that the agents, either the controlled or the enemies, share the same state space S' of the current game; and within each camp, agents are homogeneousâ thus having the same action spaces A and B respectively. That is, for each agent i ⬠[1, N] and enemy j ⬠[1, M], Ai = A and B; = B. As discrete action space is intractably large, we consider continuous control outputs, e.g., attack angle and distance.
In defining the reward function, we introduce a time- variant global reward based on the difference of the heath level between two consecutive time steps:
N+M + 1 N r(s,a,b) = M > ARâ(s,a,b) - WD ARi(s.a,b), j=N4+1 i=1 (1) | 1703.10069#12 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 12 | Spectral analysis Character embedding Encoder CBHG Encoder pre-net Decoder pre-net Decoder RNN Attention RNN Post-processing net CBHG pre-emphasis: 0.97; frame length: 50 ms; frame shift: 12.5 ms; window type: Hann 256-D Conv1D bank: K=16, conv-k-128-ReLU Max pooling: stride=1, width=2 Conv1D projections: conv-3-128-ReLU â conv-3-128-Linear Highway net: 4 layers of FC-128-ReLU Bidirectional GRU: 128 cells FC-256-ReLU â Dropout(0.5) â FC-128-ReLU â Dropout(0.5) FC-256-ReLU â Dropout(0.5)â FC-128-ReLU â Dropout(0.5) 2-layer residual GRU (256 cells) 1-layer GRU (256 cells) Conv1D bank: K=8, conv-k-128-ReLU Max pooling: stride=1, width=2 Conv1D projections: conv-3-256-ReLU â conv-3-80-Linear Highway net: 4 layers of FC-128-ReLU Bidirectional GRU: 128 cells 2 Reduction factor (r) | 1703.10135#12 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 13 | not explicitly optimized for computation efï¬ciency and consequently our experiments show that MSDNets substantially outperform FractalNets. Our dynamic evaluation strategy for reducing batch computational cost is closely related to the the adaptive computation time approach (Graves, 2016; Figurnov et al., 2016), and the recently proposed method of adaptively evaluating neural networks (Bolukbasi et al., 2017). Different from these works, our method adopts a specially designed net- work with multiple classiï¬ers, which are jointly optimized during training and can directly output conï¬dence scores to control the evaluation process for each test example. The adaptive computation time method (Graves, 2016) and its extension (Figurnov et al., 2016) also perform adaptive eval- uation on test examples to save batch computational cost, but focus on skipping units rather than layers. In (Odena et al., 2017), a âcomposerâmodel is trained to construct the evaluation network from a set of sub-modules for each test example. By contrast, our work uses a single CNN with multiple intermediate classiï¬ers that is trained end-to-end. The Feedback | 1703.09844#13 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 13 | where for simplicity, we drop the subscript ¢ in global reward r(s,a, b). For given time step ¢ with state s, the controlled agents take actions a ⬠AY, the opponents take actions b ⬠Bâ¢, and AR*(.) = Ri7!(s,a,b) â Ri(s, a, b) repre- sents the reduced health level for agent 7. Note that Eq.(1) is presented from the aspect of controlled agents; the enemyâs global reward is the exact opposite, making the sum of re- wards from both camps equal zero. As the health level is non-increasing over time, Eq. (1) gives a positive reward at time step ¢ if the decrease of enemiesâ health levels exceeds that of ours.
With the defined global reward r(s, a, b), the controlled agents jointly take actions a in state s when the enemies take joint actions b. The agentsâ objective is to learn a policy that maximises the expected sum of discounted rewards, i.e., Ets AM ritn], where 0 < \ < 1 is discount factor. Con- versely, the enemiesâ joint policy is to minimise the expected sum. Correspondingly, we have the following Minimax game: | 1703.10069#13 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 13 | bedded into a continuous vector. We then apply a set of non-linear transformations, collectively called a âpre-netâ, to each embedding. We use a bottleneck layer with dropout as the pre-net in this work, which helps convergence and improves generalization. A CBHG module transforms the pre- net outputs into the ï¬nal encoder representation used by the attention module. We found that this CBHG-based encoder not only reduces overï¬tting, but also makes fewer mispronunciations than a standard multi-layer RNN encoder (see our linked page of audio samples).
# 3.3 DECODER | 1703.10135#13 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.10069 | 14 | Qiq(s, a, b) =r(s,a, b) + Amaxmin Qjq(s', aa(s"), ba(s')), (2)
where sâ = sâ+! is determined by T(s, a, b). Q§,(s, a, b) is the optimal action-state value function, which follows the Bellman Optimal Equation. Here we propose to use deter- ministic policy ag : S + AY of the controlled agents and the deterministic policy (Silver et al. 2014) bg : S + BY
With our framework heterogeneous agents can be also trained using different parameters and action space.
of the enemies. In small-scale MARL problems, a common solution is to employ Minimax Q-learning (Littman 1994). However, minimax Q-learning is generally intractable to ap- ply in complex games. For simplicity, we consider the case that the policy of enemies is fixed, while leaving dedicated opponent modelling for future work. Then, SG defined in Eq. (2) effectively turns into an MDP problem (He et al. 2016):
Q"*(s,a) =r(s,a) + Amax Q"(s',a9(sâ)), GB)
where we drop notation bg for brevity.
# Local, Individual Rewards | 1703.10069#14 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 14 | # 3.3 DECODER
We use a content-based tanh attention decoder (see e.g. Vinyals et al. (2015)), where a stateful recur- rent layer produces the attention query at each decoder time step. We concatenate the context vector and the attention RNN cell output to form the input to the decoder RNNs. We use a stack of GRUs with vertical residual connections (Wu et al., 2016) for the decoder. We found the residual con- nections speed up convergence. The decoder target is an important design choice. While we could directly predict raw spectrogram, itâs a highly redundant representation for the purpose of learning alignment between speech signal and text (which is really the motivation of using seq2seq for this task). Because of this redundancy, we use a different target for seq2seq decoding and waveform syn- thesis. The seq2seq target can be highly compressed as long as it provides sufï¬cient intelligibility and prosody information for an inversion process, which could be ï¬xed or trained. We use 80-band mel-scale spectrogram as the target, though fewer bands or more concise targets such as cepstrum could be used. We use a post-processing network (discussed below) to convert from the seq2seq target to waveform. | 1703.10135#14 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 15 | Related network architectures. Our network architecture borrows elements from neural fabrics (Saxena & Verbeek, 2016) and others (Zhou et al., 2015; Jacobsen et al., 2017; Ke et al., 2016)
3
Published as a conference paper at ICLR 2018
Relative accuracy of the intermediate classifier Relative accuracy of the final classifier Lo} <p he 1.00 ,O9F uc 4 0.98 4 Bos . â f z 0.87 ? â 4 4 L , = 0.96 4 â , é OTF 2 4 ry â fal Z 0.94 4 4 S 0.6} = + . 0.92 © MSDNet (with intermediate classifier) |7 ost H © DenseNet (with intermediate classifier) 0.90 @â® ResNet (with intermediate classifier) [4 0.0 02 04 06 0.8 10 0.0 02 04 0.6 08 10 location of intermediate classifier (relative to full depth) location of intermediate classifier (relative to full depth)
# A 5 a
# oe 5
# S So | 1703.09844#15 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 15 | where we drop notation bg for brevity.
# Local, Individual Rewards
A potential drawback of only using the global reward in Eq. (1) and its resulting zero-sum game is that it ignores the fact that a team collaboration typically consists of local col- laborations and reward function and would normally includes certain internal structure. Moreover, in practice, each agent tends to have its own objective which drives the collaboration. To model this, we extend the formulation in the previous sec- tion by replacing Eq. (1) with each agentâs local reward and including the evaluation of their attribution to other agents that have been interacting with (either completing or collabo- rating), i.e.,
= 1 t rj(s,a,b) = jtop-K-u()| > AR;(s, a, b)â Jâ¬top-K-u(i) 1 t Toke 2 ARH ab), @ i! â¬top-K-e(i)
where each agent i maintains top-K-u(i) and top-K-e(i), the top-C lists of other agents and enemies, that are cur- rently being interacted with. Replacing it with Eq. (1), we have N number of Bellman equations for agent i, where i ⬠{1,..., N}, for the same parameter 0 of the policy func- tion: | 1703.10069#15 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 15 | We use a simple fully-connected output layer to predict the decoder targets. An important trick we discovered was predicting multiple, non-overlapping output frames at each decoder step. Predicting r frames at once divides the total number of decoder steps by r, which reduces model size, training time, and inference time. More importantly, we found this trick to substantially increase convergence speed, as measured by a much faster (and more stable) alignment learned from attention. This is likely because neighboring speech frames are correlated and each character usually corresponds to multiple frames. Emitting one frame at a time forces the model to attend to the same input token for multiple timesteps; emitting multiple frames allows the attention to move forward early in training. A similar trick is also used in Zen et al. (2016) but mainly to speed up inference.
4 | 1703.10135#15 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.10069 | 16 | Qi(s,a) =ri(s,a) + AmaxQj(s',ag(sâ)). â )
# Communication w/ Bidirectional Backpropagation
Although Eq. (5) makes single-agent methods, such as deter- ministic policy gradient (Silver et al. 2014; Mnih et al. 2016), immediately applicable for learning individual actions, those approaches, however, lacks a principled mechanism to foster team-level collaboration. In this paper, we allow communica- tions between agents (right before taking individual actions) by proposing a bidirectionally-coordinated net (BiCNet).
Overall, BiCNet consists of a multiagent actor network and a multiagent critic network as illustrated in Fig.(1). Both of the policy network (actor) and the Q-network (critic) are based on the bi-directional RNN structure (Schuster and Pali- wal 1997). The policy network, which takes in a shared ob- servation together with a local view, returns the action for each individual agent. As the bi-directional recurrent struc- ture could serve not only as a communication channel but also as a local memory saver, each individual agent is able to maintain its own internal states, as well as to share the information with its collaborators. | 1703.10069#16 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 16 | 4
The ï¬rst decoder step is conditioned on an all-zero frame, which represents a <GO> frame. In inference, at decoder step t, the last frame of the r predictions is fed as input to the decoder at step t + 1. Note that feeding the last prediction is an ad-hoc choice here â we could use all r predictions. During training, we always feed every r-th ground truth frame to the decoder. The input frame is passed to a pre-net as is done in the encoder. Since we do not use techniques such as scheduled sampling (Bengio et al., 2015) (we found it to hurt audio quality), the dropout in the pre-net is critical for the model to generalize, as it provides a noise source to resolve the multiple modalities in the output distribution.
# 3.4 POST-PROCESSING NET AND WAVEFORM SYNTHESIS | 1703.10135#16 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 17 | to rapidly construct a low-resolution feature map that is amenable to classiï¬cation, whilst also maintaining feature maps of higher resolution that are essential for obtaining high classiï¬cation accuracy. Our design differs from the neural fabrics (Saxena & Verbeek, 2016) substantially in that MSDNets have a reduced number of scales and no sparse channel connectivity or up-sampling paths. MSDNets are at least one order of magnitude more efï¬cient and typically more accurate â for example, an MSDNet with less than 1 million parameters obtains a test error below 7.0% on CIFAR-10 (Krizhevsky & Hinton, 2009), whereas Saxena & Verbeek (2016) report 7.43% with over 20 million parameters. We use the same feature-concatenation approach as DenseNets (Huang et al., 2017), which allows us to bypass features optimized for early classiï¬ers in later layers of the network. Our architecture is related to deeply supervised networks (Lee et al., 2015) in that it incorporates classiï¬ers at multiple layers throughout the network. In contrast to all these prior architectures, our network is speciï¬cally designed to operate in resource-aware settings.
# 3 PROBLEM SETUP
We consider two settings that impose computational constraints at prediction time. | 1703.09844#17 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 17 | For the learning over BiCNet, intuitively, we can think of computing the backward gradients by unfolding the network of length N (the number of controlled agents) and then ap- plying backpropagation through time (BPTT) (Werbos 1990).
The gradients pass to both the individual Q,; function and the policy function. They are aggregated from all the agents and their actions. In other words, the gradients from all agents rewards are first propagated to influence each of agents ac- tions, and the resulting gradients are further propagated back to updating the parameters.
To see this mathematically, we denote the objective of a single agent i by J; (0); that is to maximise its expected cumu- lative individual reward r; as J;(9) = Eswpr [ri(s, ao(s))], ag where pL, (s) is the discounted state distribution correspond- ing to the policy ag under the transition T, i.e., pZ,(s) := Js De A *pi(s)1 (8! = Tey, py, (S))4s 5 it can also be cho- sen as the stationary distribution of an ergodic MDP. So, we can write the objective of N agents denoted by J(@) as follows: | 1703.10069#17 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 17 | # 3.4 POST-PROCESSING NET AND WAVEFORM SYNTHESIS
As mentioned above, the post-processing netâs task is to convert the seq2seq target to a target that can be synthesized into waveforms. Since we use Grifï¬n-Lim as the synthesizer, the post-processing net learns to predict spectral magnitude sampled on a linear-frequency scale. Another motivation of the post-processing net is that it can see the full decoded sequence. In contrast to seq2seq, which always runs from left to right, it has both forward and backward information to correct the prediction error for each individual frame. In this work, we use a CBHG module for the post-processing net, though a simpler architecture likely works as well. The concept of a post-processing network is highly general. It could be used to predict alternative targets such as vocoder parameters, or as a WaveNet-like neural vocoder (van den Oord et al., 2016; Mehri et al., 2016; Arik et al., 2017) that synthesizes waveform samples directly. | 1703.10135#17 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 18 | # 3 PROBLEM SETUP
We consider two settings that impose computational constraints at prediction time.
Anytime prediction. In the anytime prediction setting (Grubb & Bagnell, 2012), there is a ï¬nite computational budget B > 0 available for each test example x. The computational budget is nonde- terministic, and varies per test instance. It is determined by the occurrence of an event that requires the model to output a prediction immediately. We assume that the budget is drawn from some joint distribution P (x, B). In some applications P (B) may be independent of P (x) and can be estimated. For example, if the event is governed by a Poisson process, P (B) is an exponential distribution. We denote the loss of a model f (x) that has to produce a prediction for instance x within budget B by L(f (x), B). The goal of an anytime learner is to minimize the expected loss under the budget dis- tribution: L(f ) = E [L(f (x), B)]P (x,B). Here, L( ) denotes a suitable loss function. As is common · in the empirical risk minimization framework, the expectation under P (x, B) may be estimated by an average over samples from P (x, B). | 1703.09844#18 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 18 | N J(8) =Eswor [D> ri(s,ae(s))]- 6) i=1
Next, we introduce a multiagent analogue to the deterministic policy gradient theorem. The proof, which we give in the Supplementary Material, follows a similar scheme to both (Silver et al. 2014) and (Sutton et al. 2000).
Theorem 1 (Multiagent Deterministic PG Theorem) Given N agents which are collectively represented in a policy parameterised with 0, the discounted state distribution pi, (s), and the objective function J(0) defined in Eq.(6), we have the policy gradient as follows:
VoJ(0) =
N N Es pF, (s) > > Voaj,(s) - Va;Q;*(s, aa(s)) i=l j=l (7) | 1703.10069#18 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 18 | We use the Grifï¬n-Lim algorithm (Grifï¬n & Lim, 1984) to synthesize waveform from the predicted spectrogram. We found that raising the predicted magnitudes by a power of 1.2 before feeding to Grifï¬n-Lim reduces artifacts, likely due to its harmonic enhancement effect. We observed that Grifï¬n-Lim converges after 50 iterations (in fact, about 30 iterations seems to be enough), which is reasonably fast. We implemented Grifï¬n-Lim in TensorFlow (Abadi et al., 2016) hence itâs also part of the model. While Grifï¬n-Lim is differentiable (it does not have trainable weights), we do not impose any loss on it in this work. We emphasize that our choice of Grifï¬n-Lim is for simplicity; while it already yields strong results, developing a fast and high-quality trainable spectrogram to waveform inverter is ongoing work.
# 4 MODEL DETAILS
Table 1 lists the hyper-parameters and network architectures. We use log magnitude spectrogram with Hann windowing, 50 ms frame length, 12.5 ms frame shift, and 2048-point Fourier transform. We also found pre-emphasis (0.97) to be helpful. We use 24 kHz sampling rate for all experiments. | 1703.10135#18 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 19 | Budgeted batch classiï¬cation. classify a set of examples x1, . . . , xM } is known in advance. The learner aims to minimize the loss across all examples in cumulative cost bounded by B, which we denote by L(f ( ). It can potentially do so by spending less than B L( · whilst using more than B B considered here is a soft constraint when we have a large batch of testing samples.
# 4 MULTI-SCALE DENSE CONVOLUTIONAL NETWORKS
A straightforward solution to the two problems introduced in Section 3 is to train multiple networks of increasing capacity, and sequentially evaluate them at test time (as in Bolukbasi et al. (2017)). In the anytime setting the evaluation can be stopped at any point and the most recent prediction is returned. In the batch setting, the evaluation is stopped prematurely the moment a network classiï¬es
4
Published as a conference paper at ICLR 2018 | 1703.09844#19 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 19 | N N Es pF, (s) > > Voaj,(s) - Va;Q;*(s, aa(s)) i=l j=l (7)
where to ensure adequate exploration, we apply Ornstein- Uhlenbeck process to add noise on the output of actor net- work in each time step. Here we further consider the off- policy deterministic actor-critic algorithms (Lillicrap et al. 2015; Silver et al. 2014) to reduce the variance. In particular, we employ a critic function in Eq. (7) to estimate the action- value Q?? where off-policy explorations can be conducted. In training the critic, we use the sum of square loss and have the following gradient for the parametrised critic Q§(s, a), where ⬠is the parameter for the Q network:
N VeL(S) = Borys c)| Do(rl5sa0(s)) + AQE!-a0(s') i=1 ~Q§(s, av(s))) Vac (s.a0(8) . (8) | 1703.10069#19 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 19 | We use r = 2 (output layer reduction factor) for the MOS results in this paper, though larger r values (e.g. r = 5) also work well. We use the Adam optimizer with learning rate decay, which starts from 0.001 and is reduced to 0.0005, 0.0003, and 0.0001 after SOOK, 1M and 2M global steps, respectively. We use a simple ¢1 loss for both seq2seq decoder (mel-scale spectrogram) and post-processing net (linear-scale spectrogram). The two losses have equal weights.
We train using a batch size of 32, where all sequences are padded to a max length. Itâs a com- mon practice to train sequence models with a loss mask, which masks loss on zero-padded frames. However, we found that models trained this way donât know when to stop emitting outputs, causing repeated sounds towards the end. One simple trick to get around this problem is to also reconstruct the zero-padded frames.
# 5 EXPERIMENTS
We train Tacotron on an internal North American English dataset, which contains about 24.6 hours of speech data spoken by a professional female speaker. The phrases are text normalized, e.g. â16â is converted to âsixteenâ.
5 | 1703.10135#19 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 20 | 4
Published as a conference paper at ICLR 2018
the test sample with sufï¬cient conï¬dence. When the resources are so limited that the execution is terminated after the ï¬rst network, this approach is optimal because the ï¬rst network is trained for exactly this computational budget without compromises. However, in both settings, this scenario is rare. In the more common scenario where some test samples can require more processing time than others the approach is far from optimal because previously learned features are never re-used across the different networks.
An alternative solution is to build a deep network with a cascade of classiï¬ers operating on the features of internal layers: in such a network features computed for an earlier classiï¬er can be re-used by later classiï¬ers. However, na¨ıvely attaching intermediate early-exit classiï¬ers to a state- of-the-art deep network leads to poor performance. | 1703.09844#20 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 20 | Note that the gradient is also aggregated from multiple agents as the policy network would do. With Eqs. (7) and Eqs. (8), we apply Stochastic Gradient Descent (SGD) to op- timise both the actor and the critic networks. The pseudocode of the over algorithm is given in the Supplementary Material.
BiCNet is markedly different from greedy MDP approach as the dependency of agents are embedded in the latent lay- ers, rather than directly on the actions. While simple, our
approach allow full dependency among agents because the gradients from all the actions in Eq.(7) are efficiently prop- agated through the entire networks. Yet, unlike CommNet (Sukhbaatar, Fergus, and others 2016), our communication is not fully symmetric, and we maintain certain social conven- tions and roles by fixing the order of the agents that join the RNN. This would help solving any possible tie between multi- ple optimal joint actions (Busoniu, Babuska, and De Schutter 2008; Spaan et al. 2002). | 1703.10069#20 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 20 | 5
Encoder states 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 0 50 100 150 200 250 300 350 Decoder timesteps
(a) Vanilla seq2seq + scheduled sampling
0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 Encoder states 0 10 20 30 40 50 60 70 Decoder timesteps
(b) GRU encoder
0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 Encoder states 0 10 20 30 40 50 60 70 Decoder timesteps
(c) Tacotron (proposed)
Figure 3: Attention alignments on a test phrase. The decoder length in Tacotron is shorter due to the use of the output reduction factor r=5.
5.1 ABLATION ANALYSIS
We conduct a few ablation studies to understand the key components in our model. As is common for generative models, itâs hard to compare models based on objective metrics, which often do not correlate well with perception (Theis et al., 2015). We mainly rely on visual comparisons instead. We strongly encourage readers to listen to the provided samples. | 1703.10135#20 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 21 | There are two reasons why intermediate early-exit classiï¬ers hurt the performance of deep neural networks: early classiï¬ers lack coarse-level features and classiï¬ers throughout interfere with the feature generation process. In this section we investigate these effects empirically (see Figure 3) and, in response to our ï¬ndings, propose the MSDNet architecture illustrated in Figure 2. Problem: The lack of coarse-level features. Traditional neural networks learn features of ï¬ne scale in early layers and coarse scale in later layers (through repeated convolution, pooling, and strided convolution). Coarse scale features in the ï¬nal layers are important to classify the content of the whole image into a single class. Early layers lack coarse-level features and early-exit clas- siï¬ers attached to these layers will likely yield unsatisfactory high error rates. To illustrate this point, we attached4 intermediate classiï¬ers to varying layers of a ResNet (He et al., 2016) and a DenseNet (Huang et al., 2017) on the CIFAR-100 dataset (Krizhevsky & Hinton, 2009). The blue | 1703.09844#21 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 21 | Across different agents, the parameters are shared so that the number of parameters is independent of the number of agents (analogous to the shared parameters across the time domain in vanilla RNN). Parameter sharing results in the compactness of the model which could speed up the learning process. Moreover, this could also enable the domain adap- tion where the network trained on the small team of of agents (typically three) effectively scales up to larger team of agents during the test under different combat scenarios.
# Experiments
# Experimental Setup
To understand the properties of our proposed BiCNet and its performance, we conducted the experiments on the Star- Craft combats with different settings . Following similar ex- periment set-up as Sukhbaatar, Fergus, and others, BiCNet controls a group of agents trying to defeat the enemy units controlled by the built-in AI. | 1703.10069#21 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 21 | First, we compare with a vanilla seq2seq model. Both the encoder and decoder use 2 layers of residual RNNs, where each layer has 256 GRU cells (we tried LSTM and got similar results). No pre-net or post-processing net is used, and the decoder directly predicts linear-scale log magnitude spectrogram. We found that scheduled sampling (sampling rate 0.5) is required for this model to learn alignments and generalize. We show the learned attention alignment in Figure 3. Figure 3(a) reveals that the vanilla seq2seq learns a poor alignment. One problem is that attention tends to
6
1000 800 DFT bin 400 200 Frame
(a) Without post-processing net
1000 800 600 DFT bin 400 200 Frame
# (b) With post-processing net
Figure 4: Predicted spectrograms with and without using the post-processing net.
get stuck for many frames before moving forward, which causes bad speech intelligibility in the synthesized signal. The naturalness and overall duration are destroyed as a result. In contrast, our model learns a clean and smooth alignment, as shown in Figure 3(c). | 1703.10135#21 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 22 | 2016) and a DenseNet (Huang et al., 2017) on the CIFAR-100 dataset (Krizhevsky & Hinton, 2009). The blue and red dashed lines in the left plot of Figure 3 show the relative accuracies of these classiï¬ers. All three plots gives rise to a clear trend: the accuracy of a classiï¬er is highly correlated with its position within the network. Particularly in the case of the ResNet (blue line), one can observe a visible âstaircaseâ pattern, with big improvements after the 2nd and 4th classiï¬ers â located right after pooling layers. | 1703.09844#22 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 22 | The level of combat difficulties can be adjusted by vary- ing the unit types and the number of units in both sides. We measured the winning rates, and compared it with the state-of- the-art approaches. The comparative baselines consist of both the rule-based approaches, and deep reinforcement learning approaches. Our setting is summarised as follows where BiC- Net controls the former units and the built-in AI controls the latter. We categorize the settings into three types: 1) easy combats {3 Marines vs. 1 Super Zergling, and 3 Wraiths vs. 3 Mutalisks}; 2) hard combats {5 Marines vs. 5 Marines, 15 Marines vs. 16 Marines, 20 Marines vs. 30 Zerglings, 10 Marines vs. 13 Zerglings, and 15 Wraiths vs. 17 Wraiths.}; 3) heterogeneous combats { 2 Dropships and 2 Tanks vs. 1 Ultralisk }.
The rule-based approaches allow us to have a reference point that we could make sense of. Here we adopted three rule-based baselines: StarCraft built-in AI, Attack the Weakest, Attack the Closest.
For the deep reinforcement learning approaches, we con- sidered the following work as the baselines: | 1703.10069#22 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 22 | Second, we compare with a model with the CBHG encoder replaced by a 2-layer residual GRU encoder. The rest of the model, including the encoder pre-net, remain exactly the same. Comparing Figure 3(b) and 3(c), we can see that the alignment from the GRU encoder is noisier. Listening to synthesized signals, we found that noisy alignment often leads to mispronunciations. The CBHG encoder reduces overï¬tting and generalizes well to long and complex phrases.
Figures 4(a) and 4(b) demonstrate the beneï¬t of using the post-processing net. We trained a model without the post-processing net while keeping all the other components untouched (except that the decoder RNN predicts linear-scale spectrogram). With more contextual information, the prediction from the post-processing net contains better resolved harmonics (e.g. higher harmonics between bins 100 and 400) and high frequency formant structure, which reduces synthesis artifacts.
# 5.2 MEAN OPINION SCORE TESTS
We conduct mean opinion score tests, where the subjects were asked to rate the naturalness of the stimuli in a 5-point Likert scale score. The MOS tests were crowdsourced from native speakers.
7 | 1703.10135#22 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 23 | Solution: Multi-scale feature maps. To address this issue, MSDNets maintain a feature repre- sentation at multiple scales throughout the network, and all the classiï¬ers only use the coarse-level features. The feature maps at a particular layer5 and scale are computed by concatenating the re- sults of one or two convolutions: 1. the result of a regular convolution applied on the same-scale features from the previous layer (horizontal connections) and, if possible, 2. the result of a strided convolution applied on the ï¬ner-scale feature map from the previous layer (diagonal connections). The horizontal connections preserve and progress high-resolution information, which facilitates the construction of high-quality coarse features in later layers. The vertical connections produce coarse features throughout that are amenable to classiï¬cation. The dashed black line in Figure 3 shows that MSDNets substantially increase the accuracy of early classiï¬ers. Problem: Early classiï¬ers interfere with later classiï¬ers. The right plot of Figure 3 shows the accuracies of the ï¬nal classiï¬er as a function of the location of a single intermediate | 1703.09844#23 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 23 | For the deep reinforcement learning approaches, we con- sidered the following work as the baselines:
Independent controller (IND): We trained the model for single agent and control each agent individually in the com- bats. Note that there is no information sharing among differ- ent agents even though such method is easily adaptable to all kinds of multiagent combats.
Fully-connected (FC): We trained the model for all agents in a multiagent setting and control them collectively in the com- bats. The communication between agents are fully-connected. Note that it is needed to re-train a different model when the number of units at either side changes.
CommNet: CommNet (Sukhbaatar, Fergus, and others 2016) is a multiagent network designed to learning to communicate among multiple agents. To make a fair comparison, we im- plemented both the CommNet and the BiCNet on the same (state, action) spaces and follow the same training processes.
âmem batch size 16 um batch sive 32 mmm batch size 64 mm batch size 128 y = Winning Rate y =MeanQ 100 ale â lo Training Testing 200 300 400-=«500. «600700 x= Dataset X= Max Number of Episodes
Figure 2: The impact of batch_size in combat 2 Marines vs. 1 Super Zergling. | 1703.10069#23 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 23 | 7
100 unseen phrases were used for the tests and each phrase received 8 ratings. When computing MOS, we only include ratings where headphones were used. We compare our model with a para- metric (based on LSTM (Zen et al., 2016)) and a concatenative system (Gonzalvo et al., 2016), both of which are in production. As shown in Table 2, Tacotron achieves an MOS of 3.82, which outperforms the parametric system. Given the strong baselines and the artifacts introduced by the Grifï¬n-Lim synthesis, this represents a very promising result.
# Table 2: 5-scale mean opinion score evaluation.
Tacotron Parametric Concatenative mean opinion score 3.82 ± 0.085 3.69 ± 0.109 4.09 ± 0.119
# 6 DISCUSSIONS | 1703.10135#23 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 24 | The right plot of Figure 3 shows the accuracies of the ï¬nal classiï¬er as a function of the location of a single intermediate classiï¬er, relative to the accuracy of a network without intermediate classiï¬ers. The results show that the introduction of an intermediate classiï¬er harms the ï¬nal ResNet classiï¬er (blue line), reducing its accuracy by up to 7%. We postulate that this accuracy degradation in the ResNet may be caused by the intermediate classiï¬er inï¬uencing the early features to be optimized for the short-term and not for the ï¬nal layers. This improves the accuracy of the immediate classiï¬er but collapses information required to generate high quality features in later layers. This effect becomes more pronounced when the ï¬rst classiï¬er is attached to an earlier layer. | 1703.09844#24 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 24 | Figure 2: The impact of batch_size in combat 2 Marines vs. 1 Super Zergling.
GMEZO: GreedyMDP with Episodic Zero-Order Optimisa- tion (GMEZO) (Usunier et al. 2016) was proposed particu- larly to solve StarCraft micromanagement tasks. Two novel ideas are introduced: conducting collaborations through a greedy update over MDP agents, as well as adding episodic noises in the parameter space for explorations. To focus on the comparison with these two ideas, we replaced our bi- directional formulation with the greedy MDP approach, and employed episodic zero-order optimisation with noise over the parameter space in the last layer of Q networks in our BiCNet. We keep the rest of the settings exactly the same. | 1703.10069#24 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 24 | # 6 DISCUSSIONS
We have proposed Tacotron, an integrated end-to-end generative TTS model that takes a character sequence as input and outputs the corresponding spectrogram. With a very simple waveform syn- thesis module, it achieves a 3.82 MOS score on US English, outperforming a production parametric system in terms of naturalness. Tacotron is frame-based, so the inference is substantially faster than sample-level autoregressive methods. Unlike previous work, Tacotron does not need hand- engineered linguistic features or complex components such as an HMM aligner. It can be trained from scratch with random initialization. We perform simple text normalization, though recent ad- vancements in learned text normalization (Sproat & Jaitly, 2016) may render this unnecessary in the future.
We have yet to investigate many aspects of our model; many early design decisions have gone unchanged. Our output layer, attention module, loss function, and Grifï¬n-Lim-based waveform synthesizer are all ripe for improvement. For example, itâs well known that Grifï¬n-Lim outputs may have audible artifacts. We are currently working on fast and high-quality neural-network-based spectrogram inversion.
ACKNOWLEDGMENTS
The authors would like to thank Heiga Zen and Ziang Xie for constructive discussions and feedback.
# REFERENCES | 1703.10135#24 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 25 | Solution: Dense connectivity. By contrast, the DenseNet (red line) suffers much less from this effect. Dense connectivity (Huang et al., 2017) connects each layer with all subsequent layers and allows later layers to bypass features optimized for the short-term, to maintain the high accuracy of the ï¬nal classiï¬er. If an earlier layer collapses information to generate short-term features, the lost information can be recovered through the direct connection to its preceding layer. The ï¬nal classiï¬erâs performance becomes (more or less) independent of the location of the intermediate
4We select six evenly spaced locations for each of the networks to introduce the intermediate classiï¬er. Both the ResNet and DenseNet have three resolution blocks; each block offers two tentative locations for the intermediate classiï¬er. The loss of the intermediate and ï¬nal classiï¬ers are equally weighted.
5Here, we use the term âlayerâ to refer to a column in Figure 2.
5
Published as a conference paper at ICLR 2018
or directly indirectly not ze f=1 0=2 (=3 t=4 connected connected connected | 1703.09844#25 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 25 | BiCNet: In BiCNet, we defined the action space differently from Sukhbaatar, Fergus, and others. Specifically, the ac- tion space of each individual agent is represented as a 3- dimensional real vector, i.e., continuous action space. The first dimension corresponds to the probability of attack, ac- cording to which we sample a value from [0,1]. If the sampled value is 1, then the agent attacks; otherwise, the agent moves. The second and the third dimension correspond to the degree and the distance of where to attack. With the above three quantities, BiCNet can precisely order an agent to attack a certain location. Note that this is different from executing high-level commands such as âattack enemy_idâ, in other words, how to effectively output the damage is itself a form of intelligence.
# Parameter Tuning
In our training, Adam (Kingma and Ba 2014) is set as the optimiser with learning rate equal to 0.002 and the other arguments set by default values. We set the maximum steps of each episode as 800. | 1703.10069#25 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 25 | ACKNOWLEDGMENTS
The authors would like to thank Heiga Zen and Ziang Xie for constructive discussions and feedback.
# REFERENCES
Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. TensorFlow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467, 2016.
Yannis Agiomyrgiannakis. Vocaine the vocoder and applications in speech synthesis. In Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on, pp. 4230â 4234. IEEE, 2015.
Sercan Arik, Mike Chrzanowski, Adam Coates, Gregory Diamos, Andrew Gibiansky, Yongguo Kang, Xian Li, John Miller, Jonathan Raiman, Shubho Sengupta, and Mohammad Shoeybi. Deep voice: Real-time neural text-to-speech. arXiv preprint arXiv:1702.07825, 2017.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. | 1703.10135#25 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 26 | 5
Published as a conference paper at ICLR 2018
or directly indirectly not ze f=1 0=2 (=3 t=4 connected connected connected
Figure 4: The output x? of layer @ at the s" scale in a MSDNet. Herein, [...] denotes the concatenation operator, 7 (-) a regular convolution transformation, and h;(-) a strided convolutional. Note that the outputs of he and hj have the same feature map size; their outputs are concatenated along the channel dimension.
classiï¬er. As far as we know, this is the ï¬rst paper that discovers that dense connectivity is an important element to early-exit classiï¬ers in deep networks, and we make it an integral design choice in MSDNets.
4.1
THE MSDNET ARCHITECTURE
The MSDNet architecture is illustrated in Figure 2. We present its main components below. Addi- tional details on the architecture are presented in Appendix A. | 1703.09844#26 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 26 | We study the impact of the batch size and the results are shown in Figure 2 in the â2 Marines vs. 1 Super Zerglingâ combat. The two metrics, the winning rate and the Q value, are given. We fine-tune the batch_size by selecting the best BiCNet model which are trained on 800 episodes (more than 700k steps) and then tested on 100 independent games. The model with batch_size 32 achieves both the highest winning rate and the highest mean Q-value after 600k training steps. We also observed that skip frame 2 gave the highest mean Q-value between 300k and 600k training steps. We fix this parameter with the learned optimal value in the remaining part of our test.
In Fig. 3, we also compare the convergence speed of pa- rameter learning by plotting the winning rate against the number of training episodes. It shows that the proposed BiC- Net model has a much quicker convergence than the two main StarCraft baselines.
# Performance Comparison
Table 1 compares our proposed BiCNet model against the baselines under multiple combat scenarios. In each scenario,
Wining Rate Ss So S S eS & 8 & & & ° id â BiCNet CommNet - GMEZO e 0.0 20 40 60 80 100 120 140 160 Num. Episodes
Figure 3: Learning Curves in Combat â10 Marines vs. 13 Zerglingsâ | 1703.10069#26 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 26 | Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In Advances in Neural Information Processing Sys- tems, pp. 1171â1179, 2015.
William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. Listen, attend and spell: A neural network for large vocabulary conversational speech recognition. In Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International Conference on, pp. 4960â4964. IEEE, 2016.
8
Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.
Xavi Gonzalvo, Siamak Tazari, Chun-an Chan, Markus Becker, Alexander Gutkin, and Hanna Silen. In Proc. Inter- Recent advances in Google real-time HMM-driven unit selection synthesizer. speech, pp. 2238â2242, 2016.
Daniel Grifï¬n and Jae Lim. Signal estimation from modiï¬ed short-time fourier transform. IEEE Transactions on Acoustics, Speech, and Signal Processing, 32(2):236â243, 1984. | 1703.10135#26 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 27 | The MSDNet architecture is illustrated in Figure 2. We present its main components below. Addi- tional details on the architecture are presented in Appendix A.
First layer. The first layer (¢= 1) is unique as it includes vertical connections in Figure[2] Its main purpose is to âseedâ representations on all S scales. One could view its vertical layout as a miniature âS-layersâ convolutional network (S=3 in Figure [2p. Let us denote the output feature maps at layer 2 and scale s as x# and the original input image as x}. Feature maps at coarser scales are obtained via down-sampling. The output x} of the first layer is formally given in the top row of Figure[4] Subsequent layers. Following ), the output feature maps xj produced at subse- quent layers, ¢> 1, and scales, s, are a concatenation of transformed feature maps from all previous feature maps of scale s and s â 1 (if s > 1). Formally, the ¢-th vel of our network outputs a set of features at S scales {x}, see xP}, given in the last row of Figure|4| | 1703.09844#27 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 27 | Figure 3: Learning Curves in Combat â10 Marines vs. 13 Zerglingsâ
Table 1: Performance comparison. M: Marine, Z: Zergling, W: Wraith.
Combat Rule Based RL Based Built-in Weakest Closest /TND FC GMEZO CommNet BiCNet 20M vs. 30Z |1.00 000 870 940 .00T 880 1.00 1.00 5 Mvs.5 M}.720 900 700 310 .080 .910 950 920 15 M vs. 16 M |.610 000 670 590 440 .630 680 710 10M vs. 13 Z |.550 230 410 522 430 .570 440 640 15 W vs. 17 W}.440 000 300 310 .460 420 470 530
BiCNet is trained over 100k steps, and we measure the per- formance as the average winning rate on 100 test games. The winning rate of the built-in AI is also provided as an indicator of the level of difficulty of the combats.
As illustrated in Table 1, in 4/5 of the scenarios, BiCNet outperforms the other baseline models. In particular, when the number of agents goes beyond 10, where cohesive col- laborations are required, the margin of the performance gap between BiCNet and the second best starts to increase. | 1703.10069#27 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 27 | Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770â778, 2016.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015.
Jason Lee, Kyunghyun Cho, and Thomas Hofmann. Fully character-level neural machine translation without explicit segmentation. arXiv preprint arXiv:1610.03017, 2016.
Soroush Mehri, Kundan Kumar, Ishaan Gulrajani, Rithesh Kumar, Shubham Jain, Jose Sotelo, Aaron Courville, and Yoshua Bengio. SampleRNN: An unconditional end-to-end neural audio generation model. arXiv preprint arXiv:1612.07837, 2016. | 1703.10135#27 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 28 | Classifiers. The classifiers in MSDNets also follow the dense connectivity pattern within the coars- est scale, S, i.e., the classifier at layer @ uses all the features [x?, sey x?]. Each classifier consists of two convolutional layers, followed by one average pooling layer and one linear layer. In prac- tice, we only attach classifiers to some of the intermediate layers, and we let f,(-) denote the k⢠classifier. During testing in the anytime setting we propagate the input through the network until the budget is exhausted and output the most recent prediction. In the batch budget setting at test time, an example traverses the network and exits after classifier f), if its prediction confidence (we use the maximum value of the softmax probability as a confidence measure) exceeds a pre-determined threshold 0,. Before training, we compute the computational cost, C,, required to process the net- work up to the k" classifier. We denote by 0 <q < 1 a fixed exit probability that a sample that reaches a classifier will obtain a classification with sufficient confidence to exit. We assume that q is constant across all layers, which allows us to compute the probability that a sample exits at classifier | 1703.09844#28 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 28 | In the combat â5 M vs. 5 Mâ, where the key factor to win is to âfocus fireâ on the weak, the IND and the FC models have relatively poorer performance. We believe it is because both of the models do not come with an explicit collaboration mechanism between agents in the training stage; coordinating the attacks towards one single enemy is even challenging. On the contrary, GMEZO, CommNet, and BiCNet, which are explicitly or implicitly designed for multiagent collaboration, can grasp and master this simple strategy, thus enjoying bet- ter performances. Furthermore, it is worth mentioning that despite the second best performance on â5 Marines vs. 5 Marinesâ, our BiCNet only needs 10 combats before learn- ing the idea of âfocus fireâ, and achieves over 85% win rate, whereas CommNet needs more than 50 episodes to grasp the skill of âfocus fireâ with a much lower winning rate.
Note that the order of which side starts the first attack will influence the combat. This explains why in the combat â5 M vs. 5 Mâ, the built-in AI on the left (as the first to attack) has more advantages on the winning rate 0.720 over the built-in AI on the right, even though the number of marines at both sides is the same.
# How It Works | 1703.10069#28 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10135 | 28 | Jose Sotelo, Soroush Mehri, Kundan Kumar, JoËao Felipe Santos, Kyle Kastner, Aaron Courville, and Yoshua Bengio. Char2Wav: End-to-end speech synthesis. In ICLR2017 workshop submission, 2017.
Richard Sproat and Navdeep Jaitly. RNN approaches to text normalization: A challenge. arXiv preprint arXiv:1611.00068, 2016.
Rupesh Kumar Srivastava, Klaus Greff, and J¨urgen Schmidhuber. Highway networks. arXiv preprint arXiv:1505.00387, 2015.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pp. 3104â3112, 2014.
Paul Taylor. Text-to-speech synthesis. Cambridge university press, 2009.
Lucas Theis, A¨aron van den Oord, and Matthias Bethge. A note on the evaluation of generative models. arXiv preprint arXiv:1511.01844, 2015. | 1703.10135#28 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.09844 | 29 | classification with sufficient confidence to exit. We assume that q is constant across all layers, which allows us to compute the probability that a sample exits at classifier kas: qx = 2(1âq)*~1q, where z is a normalizing constant that ensures that )>,, p(qx) = 1. At test time, we need to ensure that the overall cost of classifying all samples in D;..,, does not exceed our budget B (in expectation). This gives rise to the constraint |Dyest| }>, dekCk < B. We can solve this constraint for g and determine the thresholds 6;, on a validation set in such a way that approximately |Dtest|qx Validation samples exit at the k" classifier. Loss functions. During training we use cross entropy loss functions L(f;,) for all classifiers and minimize a weighted cumulative loss: Bl Ucxyyed Uk WkL (fe). Herein, D denotes the training set and w; > 0 the weight of the k-th classifier. If the budget distribution P(B) is known, we can use the weights w;, to incorporate our prior knowledge about the budget B in the learning. | 1703.09844#29 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10135 | 29 | A¨aron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. WaveNet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016.
Oriol Vinyals, Åukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey Hinton. Gram- mar as a foreign language. In Advances in Neural Information Processing Systems, pp. 2773â 2781, 2015.
Wenfu Wang, Shuang Xu, and Bo Xu. First step towards end-to-end parametric TTS synthesis: Generating spectral parameters with neural attention. In Proceedings Interspeech, pp. 2243â2247, 2016.
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Googleâs neural machine trans- arXiv preprint lation system: Bridging the gap between human and machine translation. arXiv:1609.08144, 2016. | 1703.10135#29 | Tacotron: Towards End-to-End Speech Synthesis | A text-to-speech synthesis system typically consists of multiple stages, such
as a text analysis frontend, an acoustic model and an audio synthesis module.
Building these components often requires extensive domain expertise and may
contain brittle design choices. In this paper, we present Tacotron, an
end-to-end generative text-to-speech model that synthesizes speech directly
from characters. Given <text, audio> pairs, the model can be trained completely
from scratch with random initialization. We present several key techniques to
make the sequence-to-sequence framework perform well for this challenging task.
Tacotron achieves a 3.82 subjective 5-scale mean opinion score on US English,
outperforming a production parametric system in terms of naturalness. In
addition, since Tacotron generates speech at the frame level, it's
substantially faster than sample-level autoregressive methods. | http://arxiv.org/pdf/1703.10135 | Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, Yonghui Wu, Ron J. Weiss, Navdeep Jaitly, Zongheng Yang, Ying Xiao, Zhifeng Chen, Samy Bengio, Quoc Le, Yannis Agiomyrgiannakis, Rob Clark, Rif A. Saurous | cs.CL, cs.LG, cs.SD | Submitted to Interspeech 2017. v2 changed paper title to be
consistent with our conference submission (no content change other than typo
fixes) | null | cs.CL | 20170329 | 20170406 | [
{
"id": "1502.03167"
},
{
"id": "1611.00068"
},
{
"id": "1612.07837"
},
{
"id": "1603.04467"
},
{
"id": "1609.08144"
},
{
"id": "1505.00387"
},
{
"id": "1511.01844"
},
{
"id": "1609.03499"
},
{
"id": "1610.03017"
},
{
"id": "1702.07825"
}
] |
1703.10069 | 30 | Hidden states have high Q value - 100 Hidden states have low Q value Q Value Figure 4: Visualisation for 3 Marines vs. 1 Super Zergling combat. Upper Left: State with high Q value; Lower Left: State with low Q value; Right: Visualisation of hidden layer outputs for each step using TSNE, coloured by Q values. We visualise the model outputs when the coordinated cover at- tack is learned in Figure 4. The values in the last hidden layer of the critic network over 10k steps are collected and then em- beded in 2-dimensional space using t-SNE algorithm (Maaten and Hinton 2008). We observe that the steps with high Q- values are aggregated in the same area in the embedding space. For example, Figure 4 Upper Left shows that the agents attack the enemy in far distance when the enemy can- not attack the agents, and in this status, the model predicts high Q values. By contrast, in Figure 4 Lower Left, the agents suffer the damages from the enemy when it closes, which leads to low Q-values. Our next aim is to examine whether there is any semantic meaning of the information exchanged among agents be- fore their actions. However, due to the high variability of the StarCraft game, so far we have not observed any con- crete meaning yet. | 1703.10069#30 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.10069 | 31 | the information exchanged among agents be- fore their actions. However, due to the high variability of the StarCraft game, so far we have not observed any con- crete meaning yet. We instead only focus on bidirectioinal communications by considering a simpler game, where the sophistications that are not related to communications are removed. Specifically, this simpler game consists of n agents, At each round, each agent observes a randomly generated number (sampled in range [â10, 10] under truncated Gaus- sian) as its input, and nothing else. The goal for each agent is to output the sum over the inputs that all the agents observed. Each agent receives reward based on the difference between the sum and their prediction (action output). In the setting of three agents guessing the sum with one Bi-RNN communication layer (the hidden state size is 1) followed by a MLP layer, Figure 5 displays the values that have been transferred among three agents. As shown, Agent 1 passes a high value to Agent 2 when it observes a high ob- servation number. When Agent 2 communicates with Agent 3, it tends to output an âadditiveâ value between its own and previously communicated agent, i.e., agent 1. In other words, the hidden state value is increasing when the | 1703.10069#31 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 32 | 6
Published as a conference paper at ICLR 2018
the last layer of the network. One simple strategy to reduce the size of the network is by splitting it into S blocks along the depth dimension, and only keeping the coarsest (' â i + 1) scales in the iâ block (a schematic layout of this structure is shown in[Figure 9p. This reduces computational cost for both training and testing. Every time a scale is removed from the network, we add a transition layer between the two blocks that merges the concatenated features using a 1 x 1 convolution and cuts the number of channels in half before feeding the fine-scale features into the coarser scale via a strided convolution (this is similar to the DenseNet-BC architecture of|Huang et al.|(2017)). Second, since a classifier at layer ¢ only uses features from the coarsest scale, the finer feature maps in layer ¢ (and some of the finer feature maps in the previous Sâ2 layers) do not influence the prediction of that classifier. Therefore, we group the computation in âdiagonal blocksâ such that we only propagate the example along paths that are required for the evaluation of the next classifier. This minimizes unnecessary computations when we need to stop because the computational budget is exhausted. We call this strategy lazy evaluation.
# 5 EXPERIMENTS | 1703.09844#32 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 32 | to output an âadditiveâ value between its own and previously communicated agent, i.e., agent 1. In other words, the hidden state value is increasing when the sum of Agents 1 and 2âs observations goes high. Both senders have learned to make the other receiver obtain a helpful message in order to predict the target sum over all agentsâ observations. We further set the game with num. of agents n = 5, 10, or 20. Apart from the four baselines tested previously, we also implement a supervised MLP with 10 hidden nodes as additional (predicting the sum based on the inputs given to agents). The results are compared in Table 2. The metric is the absolute value of the difference between each agentâs action and target. We see our method significantly outperform others. The second best is CommNet. Possible explanation is âAgent 2 Observation Figure 5: Left: The hidden state value passed by Agent | to Agent 2 in three agent guessing number game; Middle: The hidden state value passed by Agent | and Agent 2 to Agent 3 in three agent guessing number game; Right: Colour bar. Table 2: Performance comparison in the guessing game with different agent numbers. Results are given as average laction_value â | 1703.10069#32 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 33 | We evaluate the effectiveness of our approach on three image classiï¬cation datasets, i.e., the CIFAR- 10, CIFAR-100 (Krizhevsky & Hinton, 2009) and ILSVRC 2012 (ImageNet; Deng et al. (2009)) datasets. Code to reproduce all results is available at https://anonymous-url. Details on architectural conï¬gurations of MSDNets are described in Appendix A. Datasets. The two CIFAR datasets contain 50, 000 training and 10, 000 test images of 32 32 pixels; we hold out 5, 000 training images as a validation set. The datasets comprise 10 and 100 classes, respectively. We follow He et al. (2016) and apply standard data-augmentation techniques to the training images: images are zero-padded with 4 pixels on each side, and then randomly cropped to produce 32 32 images. Images are ï¬ipped horizontally with probability 0.5, and normalized by subtracting channel means and dividing by channel standard deviations. The ImageNet dataset comprises 1, 000 classes, with a total of 1.2 million training images and 50,000 validation images. We hold out 50,000 images from the training set | 1703.09844#33 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 33 | agent guessing number game; Right: Colour bar. Table 2: Performance comparison in the guessing game with different agent numbers. Results are given as average laction_value â target_value| in 10,000 testing steps and its standard deviation; A smaller value means a better per- formance. SL-MLP is a supervised MLP as an additional baseline. t-test is conducted, and the significant ones (p-value < 0.05) compared to the second best are marked as *. gent Number SL-MLP IND CommNet GMEZO BiCNet 5 2.824238 13.92£12.0 0.57£0.4T 5.92£7.62. *0.5250.51 10 4.3143.67 15.32+13.90 1.18+0.90 9.21+8.22 *0.97+0.91 20 6.7145.31 19.67414.61 3.8843.03 13.65£11.74 *3.1242.93 that it takes an average as the message, and thus naturally fits the problem, while ours have to learn the additives implicitly. Emerged Human-level Coordination With adequate trainings from scratch, BiCNet would be able to discover several effective collaboration | 1703.10069#33 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 34 | dataset comprises 1, 000 classes, with a total of 1.2 million training images and 50,000 validation images. We hold out 50,000 images from the training set to estimate the conï¬dence threshold for classiï¬ers in MSDNet. We adopt the data augmentation scheme of He et al. (2016) at training time; at test time, we classify a 224 224 center crop of images that were resized to 256 Training Details. We train all models using the framework of Gross & Wilber (2016). On the two CIFAR datasets, all models (including all baselines) are trained using stochastic gradient descent (SGD) with mini-batch size 64. We use Nesterov momentum with a momentum weight of 0.9 without dampening, and a weight decay of 10â4. All models are trained for 300 epochs, with an initial learning rate of 0.1, which is divided by a factor 10 after 150 and 225 epochs. We apply the same optimization scheme to the ImageNet dataset, except that we increase the mini-batch size to 256, and all the models are trained for 90 epochs with learning rate drops after 30 and 60 epochs. | 1703.09844#34 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 34 | learn the additives implicitly. Emerged Human-level Coordination With adequate trainings from scratch, BiCNet would be able to discover several effective collaboration strategies. In this section, we conduct a qualitative analysis on the learned col- laboration policies from BiCNet. We refer the demonstration video to the Supplementary Material and the experimental configurations to Section Experiments. Coordinated moves without collision. We observe that, in the initial stages of learning, in Fig. 6 (a) and (b), the agents move in a rather uncoordinated way. In particular, when two agents are close to each other, one agent often unintentionally blocks the otherâs path. With the increasing rounds of train- ing (typically over 40k steps in near 50 episodes in the â3 (a) Early stage (b) Early stage (c) Well-trained (d) Well-trained of training of training Figure 6: Coordinated moves without collision in combat 3 Marines (ours) vs. 1 Super Zergling (enemy). The yellow line points out the direction each agent is going to move. | 1703.10069#34 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 36 | Baselines. There exist several baseline approaches for anytime prediction: FractalNets (Larsson et al., 2017), deeply supervised networks (Lee et al., 2015), and ensembles of deep networks of varying or identical sizes. FractalNets allow for multiple evaluation paths during inference time, which vary in computation time. In the anytime setting, paths are evaluated in order of increasing computation. In our result ï¬gures, we replicate the FractalNet results reported in the original paper (Larsson et al., 2017) for reference. Deeply supervised networks introduce multiple early-exit classi- ï¬ers throughout a network, which are applied on the features of the particular layer they are attached to. Instead of using the original model proposed in Lee et al. (2015), we use the more competitive ResNet and DenseNet architectures (referred to as DenseNet-BC in Huang et al. (2017)) as the base networks in our experiments with deeply supervised networks. We refer to these as ResNetMC and DenseNetMC, where M C stands for multiple classiï¬ers. Both networks require about 1.3 108 FLOPs when fully evaluated; the detailed | 1703.09844#36 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 36 | (a) time step 1 (b) time step 2 (c) time step 3 (d) time step 4 Figure 7: Hit and Run tactics in combat 3 Marines (ours) vs. I Zealot (enemy). (a) time step 1 (b) time step 2 (c) time step 3 (d) time step 4 Figure 8: Coordinated cover attacks in combat 4 Dragoons (ours) vs. 1 Ultralisks (enemy) Marines vs. 1 Super Zerglingâ combat setting), the number of collisions reduces dramatically. Finally, when the training be- comes stable, the coordinated moves emerge, as illustrated in Fig. 6 (c) and (d). Such coordinated moves become important in large-scale combats as shown in Fig. 9 (a) and (b). Hit and Run tactics. For human players, a common tactic of controlling agents in StarCraft combat is Hit and Run, i.e., moving the agents away if they are under attack, and fighting back again when agents stay safe. We find that BiCNet can rapidly grasp the tactic of Hit and Run, either in the case of single agent or multiple agents settings. We illustrate four consecutive movements of Hit and Run in Fig. 7. Despite the simplicity, Hit and Run serves as | 1703.10069#36 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 37 | and DenseNetMC, where M C stands for multiple classiï¬ers. Both networks require about 1.3 108 FLOPs when fully evaluated; the detailed network conï¬gurations are presented in the supplemen- tary material. In addition, we include ensembles of ResNets and DenseNets of varying or identical sizes. At test time, the networks are evaluated sequentially (in ascending order of network size) to obtain predictions for the test data. All predictions are averaged over the evaluated classiï¬ers. On | 1703.09844#37 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 37 | case of single agent or multiple agents settings. We illustrate four consecutive movements of Hit and Run in Fig. 7. Despite the simplicity, Hit and Run serves as the basis for more advanced and sophisticated collaboration tactics. Coordinated cover attack. Cover attack is a high-level collaborative strategy that is often used on the real battlefield. The essence of cover attack is to let one agent draw fire or attentions from the enemies, meanwhile, other agents take advantage of this time period or distance gap to output more harms. The difficulty of conducting cover attack lies in how to arrange the sequential moves of multiple agents in a coor- dinated hit and run way. As shown in Figs. 8, BiCNet can master it well. Starting from Fig. 8(a), BiCNet controls the bottom two Dragoons to run away from the enemy Ultralisk, while the one in the upper-right corner immediately starts to attack the enemy Ultralisk to cover them up. As a response, the enemy starts to attack the top one in time step 2. The bottom two Dragoons fight back and form another cover-up. By continuously looping this strategy over, the team of Dra- goons guarantees consecutive attack outputs to the enemy while | 1703.10069#37 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 38 | 7
Published as a conference paper at ICLR 2018
Anytime prediction on ImageNet Anytime prediction on CIFAR-100 ee MSDNet oo _ z 66 â MSDNet Ensemble of ResNets (varying depth) 50 F 64 62 tan Ensemble of Denseâ 60 . T r n a) 45 L L L 0.0 04 0.6 08 1.0 12 1d 0.0 0.2 04 0.6 1.0 12 10 14 budget (in MUL-ADD) x0) budget (in MUL-ADD) x108
Figure 5: Accuracy (top-1) of anytime prediction models as a function of computational budget on the ImageNet (left) and CIFAR-100 (right) datasets. Higher is better.
ImageNet, we compare MSDNet against a highly competitive ensemble of ResNets and DenseNets, with depth varying from 10 layers to 50 layers, and 36 layers to 121 layers, respectively.
Anytime prediction results are presented in Figure 5. The left plot shows the top-1 classiï¬cation accuracy on the ImageNet validation set. Here, for all budgets in our evaluation, the accuracy of MSDNet substantially outperforms the ResNets and DenseNets ensemble. In particular, when the 8% higher accuracy. budget ranges from 0.1
Ã
Ã
â¼
â | 1703.09844#38 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 38 | fight back and form another cover-up. By continuously looping this strategy over, the team of Dra- goons guarantees consecutive attack outputs to the enemy while minimising the team-level damages (because the en- emy wastes time in targeting different Dragoons) until the enemy is killed. Focus fire without overkill. As the number of agents in- creases, how to efficiently allocate the attacking resources becomes important. Neither scattering over all enemies nor focusing on one enemy (wasting attacking fires is also called overkill) are desired. We observe that BiCNet learns to con- trol each agent to focus their fires on particular enemies, and (a) time step 1 (b) time step 2 (c) timestep 3 (d) time step 4 Figure 9: focus fireâ in combat /5 Marines (ours) vs. 16 Marines (enemy). (a) time step 1 (b) time step 2 Figure 10: Coordinated heterogeneous agents in combat 2 Dropships and 2 tanks vs. | Ultralisk. different agents tend to move to the sides to spread the fire and avoid overkill. An example could be found in Fig.(9) Collaborations between heterogeneous agents. In Star- Craft, there are tens of types of agent units, | 1703.10069#38 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 39 | Ã
Ã
â¼
â
We evaluate more baselines on CIFAR-100 (and CIFAR-10; see supplementary materials). We observe that MSDNet substantially outperforms ResNetsMC and DenseNetsMC at any computational budget within our range. This is due to the fact that after just a few layers, MSDNets have produced low-resolution feature maps that are much more suitable for classiï¬cation than the high-resolution feature maps in the early layers of ResNets or DenseNets. MSDNet also outperforms the other baselines for nearly all computational budgets, although it performs on par with ensembles when the budget is very small. In the extremely low-budget regime, ensembles have an advantage because their predictions are performed by the ï¬rst (small) network, which is optimized exclusively for the low budget. However, the accuracy of ensembles does not increase nearly as fast when the budget is increased. The MSDNet outperforms the ensemble as soon as the latter needs to evaluate a second model: unlike MSDNets, this forces the ensemble to repeat the computation of similar low-level features repeatedly. Ensemble accuracies saturate rapidly when all networks are shallow.
5.2 BUDGETED BATCH CLASSIFICATION | 1703.09844#39 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 39 | the fire and avoid overkill. An example could be found in Fig.(9) Collaborations between heterogeneous agents. In Star- Craft, there are tens of types of agent units, each with unique functionalities, action space, strength, and weakness. For combats with different types of units involved, we expect the agents to reach win-win situations through the collaborations. In fact, heterogeneous collaborations can be easily imple- mented in our framework by limiting the parameter sharing only to the same types of the units. In this paper, we study a simple case where two Dropships and two tanks collaborate to fight against an Ultralisk. A Dropship does not have the function to attack, but it can carry maximally two ground units in the air. As shown in Fig. 10, when the Ultralisk is attacking one of the tanks, the Dropship escorts the tank to escape from the attack. At the same time, the other Dropship unloads his tank to the ground so as to attack the Ultralisk. At each side, the collaboration between the Dropship and the tank keeps iterating until the Ultralisk is destroyed. Conclusions In this paper, we have introduced a new deep multiagent re- inforcement learning. The action is learned by constructing a vectorised actor-critic | 1703.10069#39 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 40 | 5.2 BUDGETED BATCH CLASSIFICATION
In budgeted batch classiï¬cation setting, the predictive model receives a batch of M instances and a computational budget B for classifying all M instances. In this setting, we use dynamic evaluation: we perform early-exiting of âeasyâ examples at early classiï¬ers whilst propagating âhardâ examples through the entire network, using the procedure described in Section 4.
Baselines. On ImageNet, we compare the dynamically evaluated MSDNet with ï¬ve ResNets (He et al., 2016) and ï¬ve DenseNets (Huang et al., 2017), AlexNet (Krizhevsky et al., 2012), and Google- LeNet (Szegedy et al., 2015); see the supplementary material for details. We also evaluate an ensem- ble of the ï¬ve ResNets that uses exactly the same dynamic-evaluation procedure as MSDNets at test time: âeasyâ images are only propagated through the smallest ResNet-10, whereas âhardâ images are classiï¬ed by all ï¬ve ResNet models (predictions are averaged across all evaluated networks in the ensemble). We classify batches of M = 128 images. | 1703.09844#40 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 40 | is destroyed. Conclusions In this paper, we have introduced a new deep multiagent re- inforcement learning. The action is learned by constructing a vectorised actor-critic framework, where each dimension corresponds to an agent. The coordination is done by bi- directional recurrent communications in the internal layers. Through end-to-end learning, our BiCNet would be able to successfully learn several effective coordination strategies. Our experiments have demonstrated its ability to collaborate and master diverse combats in StarCraft combat games. We have also shown five human-level coordination strategies BiCNet could grasp from playing StarCraft combat games. Admittedly, quantifying the sophistication of the collabora- tions in games is challenging in general, and our analysis here is qualitative in nature. In the next step, we plan to carry on experiments of letting the machine compete with human players at different lev- els. We also plan to further investigate how the policies are communicated over the networks among agents in more com- plicated settings, and whether there is a specific language that may have emerged in StartCraft (Lazaridou, Peysakhovich, and Baroni 2016; Mordatch and Abbeel 2017). | 1703.10069#40 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 41 | On CIFAR-100, we compare MSDNet with several highly competitive baselines, including ResNets (He et al., 2016), DenseNets (Huang et al., 2017) of varying sizes, Stochastic Depth Net- works (Huang et al., 2016), Wide ResNets (Zagoruyko & Komodakis, 2016) and FractalNets (Lars- son et al., 2017). We also compare MSDNet to the ResNetMC and DenseNetMC models that were used in Section 5.1, using dynamic evaluation at test time. We denote these baselines as ResNetMC / DenseNetMC with early-exits. To prevent the result plots from becoming too cluttered, we present CIFAR-100 results with dynamically evaluated ensembles in the supplementary material. We clas- sify batches of M = 256 images at test time.
Budgeted batch classiï¬cation results on ImageNet are shown in the left panel of Figure 7. We trained three MSDNets with different depths, each of which covers a different range of compu8
Published as a conference paper at ICLR 2018 | 1703.09844#41 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 41 | # References
Brown and Sandholm 2017] Brown, N., and Sandholm, T. 2017. Safe and nested endgame solving for imperfect-information games. AAAI/AAI.
Busoniu, Babuska, and De Schutter 2008] Busoniu, L.; Babuska, R.; and De Schutter, B. 2008. A comprehensive survey of multia- gent reinforcement learning. IEEE Transactions on Systems Man and Cybernetics Part C Applications and Reviews 38(2):156.
Deboeck 1994] Deboeck, G. 1994. Trading on the edge: neural, genetic, and fuzzy systems for chaotic financial markets, volume 39. John Wiley & Sons.
Foerster et al. 2016] Foerster, J.; Assael, Y. M.; de Freitas, N.; and Whiteson, S. 2016. Learning to communicate with deep multi-agent reinforcement learning. In NJPS, 2137-2145.
Foerster et al. 2017a] Foerster, J.; Farquhar, G.; Afouras, T.; Nardelli, N.; and Whiteson, S. 2017a. Counterfactual multi-agent policy gradients. arXiv preprint arXiv:1705.08926. | 1703.10069#41 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 42 | Published as a conference paper at ICLR 2018
7 Budgeted batch classification on ImageNet Budgeted batch classification on CIFAR-100 * ResNet-H0 MSDNet with dynamic evaluation NSDNet with dynamic evaluation ensemble of Re © © MSDNet w/o dynamic evaluation sit ensemble of DenseNets Reset! with carlyenits ons â DenseNetâ¢Â© with early-exits s (He et al., 2015) lm M ResNets (He et al., 2015) x @-© DenseNets (Huang et al., 2016) al., 2016) al, 2016) 016) 0 1 2 3 4 5 00S 10 15 2.0 25 average budget (in MUL-ADD) x1? average budget (in MUL-ADD) x1?
Figure 7: Accuracy (top-1) of budgeted batch classiï¬cation models as a function of average computational budget per image the on ImageNet (left) and CIFAR-100 (right) datasets. Higher is better. | 1703.09844#42 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 42 | Foerster et al. 2017b] Foerster, J.; Nardelli, N.; Farquhar, G.; Torr, P.; Kohli, P.; Whiteson, S.; et al. 2017b. Stabilising experience replay for deep multi-agent reinforcement learning. arXiv preprint arXiv: 1702.08887.
Goertzel and Pennachin 2007] Goertzel, B., and Pennachin, C. 2007. Artificial general intelligence, volume 2. Springer.
He et al. 2016] He, H.; Boyd-Graber, J.; Kwok, K.; and Daumé III, H. 2016. Opponent modeling in deep reinforcement learning. In ICML, 1804-1813.
Kapetanakis and Kudenko 2002] Kapetanakis, S., and Kudenko, D. 2002. Reinforcement learning of coordination in cooperative multi- agent systems. AAAJ/IAAI 2002:326-331.
Kingma and Ba 2014] Kingma, D., and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv: 1412.6980.
Lauer and Riedmiller 2000] Lauer, M., and Riedmiller, M. 2000. An algorithm for distributed reinforcement learning in cooperative multi-agent systems. In JCML. | 1703.10069#42 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 43 | tational budgets. We plot the performance of each MSDNet as a gray curve; we select the best model for each budget based on its accuracy on the validation set, and plot the corresponding ac- curacy as a black curve. The plot shows that the predictions of MSDNets with dynamic evaluation are substantially more accurate than those of ResNets and DenseNets that use the same amount of 109 FLOPs, MSDNet achieves a top-1 computation. For instance, with an average budget of 1.7 6% higher than that achieved by a ResNet with the same number of accuracy of times fewer FLOPs. Compared to the computationally efï¬cient DenseNets, MSDNet uses FLOPs to achieve the same classiï¬cation accuracy. Moreover, MSDNet with dynamic evaluation allows for very precise tuning of the computational budget that is consumed, which is not possible with individual ResNet or DenseNet models. The ensemble of ResNets or DenseNets with dynamic evaluation performs on par with or worse than their individual counterparts (but they do allow for setting the computational budget very precisely). | 1703.09844#43 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 43 | Lauer and Riedmiller 2004] Lauer, M., and Riedmiller, M. 2004. Reinforcement learning for stochastic cooperative multi-agent sys- tems. In AAMA.
Lazaridou, Peysakhovich, and Baroni 2016] Lazaridou, A. Peysakhovich, A.; and Baroni, M. 2016. Multi-agent cooper- ation and the emergence of (natural) language. arXiv preprint arXiv:1612.07182.
Lillicrap et al. 2015] Lillicrap, T. P.; Hunt, J. J.; Pritzel, A.; Heess, N.; Erez, T.; Tassa, Y.; Silver, D.; and Wierstra, D. 2015. Con- tinuous control with deep reinforcement learning. arXiv preprint arXiv: 1509.02971.
Littman 1994] Littman, M. L. 1994. Markov games as a framework for multi-agent reinforcement learning. In JCML.
Lowe et al. 2017] Lowe, R.; Wu, Y.; Tamar, A.; Harb, J.; Abbeel, P.; and Mordatch, I. 2017. Multi-agent actor-critic for mixed cooperative-competitive environments. arXiv preprint arXiv:1706.02275. | 1703.10069#43 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 44 | The right panel of Figure 7 shows our results on CIFAR-100. The results show that MSDNets con- sistently outperform all baselines across all budgets. Notably, MSDNet performs on par with a 110- layer ResNet using only 1/10th of the computational budget and it is up to 5 times more efï¬cient than DenseNets, Stochastic Depth Networks, Wide ResNets, and FractalNets. Similar to results in the anytime-prediction setting, MSDNet substantially outperform ResNetsM C and DenseNetsM C with multiple intermediate classiï¬ers, which provides further evidence that the coarse features in the MSDNet are important for high performance in earlier layers. | 1703.09844#44 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 44 | Maaten and Hinton 2008] Maaten, L. v. d., and Hinton, G. 2008. Visualizing data using t-sne. Journal of Machine Learning Research 9(Nov):2579-2605.
Mnih et al. 2015] Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A. A.; Veness, J.; Bellemare, M. G.; Graves, A.; Riedmiller, M.; Fidjeland, A. K.; Ostrovski, G.; et al. 2015. Human-level control through deep reinforcement learning. Nature 518(7540):529-533.
Mnih et al. 2016] Mnih, V.; Badia, A. P.; Mirza, M.; Graves, A.; Lillicrap, T. P.; Harley, Silver, D.; and Kavukcuoglu, K. 2016. Asynchronous methods for deep reinforcement learning. In Jnterna- tional Conference on Machine Learning.
Mordatch and Abbeel 2017] Mordatch, I., and Abbeel, P. 2017. Emergence of grounded compositional language in multi-agent populations. arXiv preprint arXiv: 1703.04908.
Owen 1995] Owen, G. 1995. Game theory. Academic Press. | 1703.10069#44 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 45 | Visualization. To illustrate the ability of our ap- proach to reduce the computational requirements for classifying âeasyâ examples, we show twelve randomly sampled test images from two Ima- geNet classes in Figure 6. The top row shows âeasyâ examples that were correctly classiï¬ed and exited by the ï¬rst classiï¬er. The bottom row shows âhardâ examples that would have been in- correctly classiï¬ed by the ï¬rst classiï¬er but were passed on because its uncertainty was too high. The ï¬gure suggests that early classiï¬ers recog- nize prototypical class examples, whereas the last classiï¬er recognizes non-typical images.
_
f (a) Red wine (b) Volcano sy ws ~â
Figure 6: Sampled images from the ImageNet classes Red wine and Volcano. Top row: images exited from the ï¬rst classiï¬er of a MSDNet with correct predic- tion; Bottom row: images failed to be correctly clas- siï¬ed at the ï¬rst classiï¬er but were correctly pre- dicted and exited at the last layer.
5.3 MORE COMPUTATIONALLY EFFICIENT DENSENETS | 1703.09844#45 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 45 | Owen 1995] Owen, G. 1995. Game theory. Academic Press.
Schafer, Konstan, and Riedl 1999] Schafer, J. B.; Konstan, J.; and Riedl, J. 1999. Recommender systems in e-commerce. In ACM EC.
Schmidhuber 1996] Schmidhuber, J. 1996. A general method for multi-agent reinforcement learning in unrestricted environments. In Adaptation, Coevolution and Learning in Multiagent Systems: Papers from the 1996 AAAI Spring Symposium, 84-87.
Schuster and Paliwal 1997] Schuster, M., and Paliwal, K. K. 1997. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11):2673-2681.
Silver et al. 2014] Silver, D.; Lever, G.; Heess, N.; Degris, T.; Wier- stra, D.; and Riedmiller, M. 2014. Deterministic policy gradient algorithms. In JCML. | 1703.10069#45 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 46 | 5.3 MORE COMPUTATIONALLY EFFICIENT DENSENETS
Here, we discuss an interesting ï¬nding during our exploration of the MSDNet architecture. We found that following the DenseNet structure to design our network, i.e., by keeping the number of output channels (or growth rate) the same at all scales, did not lead to optimal results in terms of the accuracy-speed trade-off. The main reason for this is that compared to network architectures like ResNets, the DenseNet structure tends to apply more ï¬lters on the high-resolution feature maps in the network. This helps to reduce the number of parameters in the model, but at the same time, it greatly increases the computational cost. We tried to modify DenseNets by doubling the growth rate
9
Published as a conference paper at ICLR 2018
Anytime prediction on CIFAR-100 Batch computational learning on CIFAR-100 7s : : : ee â MSDNet with early-exits H 8 Del s (Huang et al., 2016) bom De st lik. . J} i: 0.0 0.2 04 0. 6 2 0.0 0.5, 10 15 2.0 2.5 budget (in MUL- ADD) x10® average budget (in MUL-ADD) x10° | 1703.09844#46 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 46 | Silver et al. 2016] Silver, D.; Huang, A.; Maddison, C. J.; Guez, A.; Sifre, L.; Van Den Driessche, G.; Schrittwieser, J.; Antonoglou, L; Panneershelvam, V.; Lanctot, M.; et al. 2016. Mastering the game of go with deep neural networks and tree search. Nature 529(7587):484489.
Spaan et al. 2002] Spaan, M. T.; Vlassis, N.; Groen, F. C.; et al. 2002. High level coordination of agents based on multiagent markov decision processes with roles. In JROS, volume 2, 66-73.
Sukhbaatar, Fergus, and others 2016] Sukhbaatar, S.; Fergus, R.; etal. 2016. Learning multiagent communication with backpropaga- tion. In NIPS, 2244-2252.
Sutton et al. 2000] Sutton, R. S.; McAllester, D. A.; Singh, S. P.; and Mansour, Y. 2000. Policy gradient methods for reinforcement learning with function approximation. In NPS, 1057-1063. | 1703.10069#46 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 47 | Figure 8: Test accuracy of DenseNet* on CIFAR-100 under the anytime learning setting (left) and the budgeted batch setting (right).
after each transition layer, so that more ï¬lters are applied to low-resolution feature maps. It turns out that the resulting network, which we denote as DenseNet*, signiï¬cantly outperform the original DenseNet in terms of computational efï¬ciency.
We experimented with DenseNet* in our two settings with test time budget constraints. The left panel of Figure 8 shows the anytime prediction performance of an ensemble of DenseNets* of vary- ing depths. It outperforms the ensemble of original DenseNets of varying depth by a large margin, but is still slightly worse than MSDNets. In the budgeted batch budget setting, DenseNet* also leads to signiï¬cantly higher accuracy over its counterpart under all budgets, but is still substantially outperformed by MSDNets.
# 6 CONCLUSION | 1703.09844#47 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 47 | Synnaeve et al. 2016] Synnaeve, G.; Nardelli, N.; Auvolat, A.; Chintala, S.; Lacroix, T.; Lin, Z.; Richoux, F.; and Usunier, N. 2016. Torchcraft: a library for machine learning research on real-time strategy games. arXiv preprint arXiv: 1611.00625.
Usunier et al. 2016] Usunier, N.; Synnaeve, G.; Lin, Z.; and Chin- tala, S. 2016. Episodic exploration for deep deterministic policies: An application to starcraft micromanagement tasks. arXiv preprint arXiv:1609.02993.
Vinyals et al. 2017] Vinyals, O.; Ewalds, T.; Bartunov, S.; Georgiev, P.; Vezhnevets, A. S.; Yeo, M.; Makhzani, A.; Kiittler, H.; Agapiou, J.; Schrittwieser, J.; et al. 2017. Starcraft ii: A new challenge for reinforcement learning. arXiv preprint arXiv: 1708.04782. | 1703.10069#47 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 48 | # 6 CONCLUSION
We presented the MSDNet, a novel convolutional network architecture, optimized to incorporate CPU budgets at test-time. Our design is based on two high-level design principles, to generate and maintain coarse level features throughout the network and to inter-connect the layers with dense connectivity. The former allows us to introduce intermediate classiï¬ers even at early layers and the latter ensures that these classiï¬ers do not interfere with each other. The ï¬nal design is a two dimensional array of horizontal and vertical layers, which decouples depth and feature coarseness. Whereas in traditional convolutional networks features only become coarser with increasing depth, the MSDNet generates features of all resolutions from the ï¬rst layer on and maintains them through- out. The result is an architecture with an unprecedented range of efï¬ciency. A single network can outperform all competitive baselines on an impressive range of computational budgets ranging from highly limited CPU constraints to almost unconstrained settings. | 1703.09844#48 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 48 | Wang, Zhang, and Yuan 2017] Wang, J.; Zhang, W.; and Yuan, S. 2017. Display advertising with real-time bidding (RTB) and be- havioural targeting. Foundations and Trends in Information Re- trieval, Now Publishers.
Werbos 1990] Werbos, P. J. 1990. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE 78(10):1550- 1560.
# Supplementary Material
# Proof of Theorem 1
# Qi" (8, a)la=ao(s)
Following the regularity conditions mentioned in (Silver et al. 2014), we know that the supreme of 0a; ,, ind
# 0aio(s)
for each agent 7 are bounded functions of s. Based on the regularity and the boundedness, we can use Leibniz integral
# c
rule and Fubiniâs theorem, respectively. Note that as the policy ag and the transition matrix of the environment 7 are both considered deterministic, the expectation is only taken over the initial state sp, which is different from the original deterministic policy gradient theorem. According to the definition of Q?°(s, a) and the our objective function in Eq.(6), we derive the multiagent deterministic policy gradient theorem, which mostly follows the line of (Sutton et al. 2000). | 1703.10069#48 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 49 | As future work we plan to investigate the use of resource-aware deep architectures beyond object classiï¬cation, e.g. image segmentation (Long et al., 2015). Further, we intend to explore approaches that combine MSDNets with model compression (Chen et al., 2015; Han et al., 2015), spatially adaptive computation (Figurnov et al., 2016) and more efï¬cient convolution operations (Chollet, 2016; Howard et al., 2017) to further improve computational efï¬ciency.
ACKNOWLEDGMENTS
The authors are supported in part by grants from the National Science Foundation ( III-1525919, IIS-1550179, IIS-1618134, S&AS 1724282, and CCF-1740822), the Ofï¬ce of Naval Research DOD (N00014-17-1-2175), and the Bill and Melinda Gates Foundation. We are also thankful for generous support by SAP America Inc.
# REFERENCES | 1703.09844#49 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.10069 | 49 | Ny âBe man JP) LO 008)) )
ad = [noZrere. ag(s))ds (10)
Ny = [m9 % (nisavis)) + [ M(s! = Te (8) "(8 as!) ) ds dh
# Ny
=
[ m6) (ee done ia} ds + [mw [a (â2 Fall = Tevs(Nlanm Do a(t) dsids + [me [aura a0.b, (8 1 eer s',a9(sâ ») ds'ds (12)
Jag(s = [ m6) 30 Fon (S, a) |aâay(s) +f L(s' = Ta», (8 ed sâ,ag(sâ)) | dsâ | ds â<$<$<â_ iterate as Eq.(10) to Eq.(11)
=
= ) Daal (sâ) a L(s' = 0 ( ndsâ 1 [LE rmente = Tes (00S goon a)jana,(s)ds'ds ( 4) | 1703.10069#49 | Multiagent Bidirectionally-Coordinated Nets: Emergence of Human-level Coordination in Learning to Play StarCraft Combat Games | Many artificial intelligence (AI) applications often require multiple
intelligent agents to work in a collaborative effort. Efficient learning for
intra-agent communication and coordination is an indispensable step towards
general AI. In this paper, we take StarCraft combat game as a case study, where
the task is to coordinate multiple agents as a team to defeat their enemies. To
maintain a scalable yet effective communication protocol, we introduce a
Multiagent Bidirectionally-Coordinated Network (BiCNet ['bIknet]) with a
vectorised extension of actor-critic formulation. We show that BiCNet can
handle different types of combats with arbitrary numbers of AI agents for both
sides. Our analysis demonstrates that without any supervisions such as human
demonstrations or labelled data, BiCNet could learn various types of advanced
coordination strategies that have been commonly used by experienced game
players. In our experiments, we evaluate our approach against multiple
baselines under different scenarios; it shows state-of-the-art performance, and
possesses potential values for large-scale real-world applications. | http://arxiv.org/pdf/1703.10069 | Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, Jun Wang | cs.AI, cs.LG | 10 pages, 10 figures. Previously as title: "Multiagent
Bidirectionally-Coordinated Nets for Learning to Play StarCraft Combat
Games", Mar 2017 | null | cs.AI | 20170329 | 20170914 | [
{
"id": "1609.02993"
},
{
"id": "1706.02275"
},
{
"id": "1705.08926"
},
{
"id": "1612.07182"
}
] |
1703.09844 | 50 | # REFERENCES
Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016.
10
Published as a conference paper at ICLR 2018
Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for fast test-time prediction. arXiv preprint arXiv:1702.07811, 2017.
Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In ACM SIGKDD, pp. 535â541. ACM, 2006.
Wenlin Chen, James T Wilson, Stephen Tyree, Kilian Q Weinberger, and Yixin Chen. Compressing neural networks with the hashing trick. In ICML, pp. 2285â2294, 2015.
Franc¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv preprint arXiv:1610.02357, 2016. | 1703.09844#50 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
1703.09844 | 51 | Franc¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv preprint arXiv:1610.02357, 2016.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pp. 248â255, 2009.
Michael Figurnov, Maxwell D Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov. Spatially adaptive computation time for residual networks. arXiv preprint arXiv:1612.02297, 2016.
Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016.
Sam Gross and Michael Wilber. Training and investigating residual nets. 2016. URL http: //torch.ch/blog/2016/02/04/resnets.html.
Alexander Grubb and Drew Bagnell. Speedboost: Anytime prediction with uniform near-optimality. In AISTATS, volume 15, pp. 458â466, 2012. | 1703.09844#51 | Multi-Scale Dense Networks for Resource Efficient Image Classification | In this paper we investigate image classification with computational resource
limits at test time. Two such settings are: 1. anytime classification, where
the network's prediction for a test example is progressively updated,
facilitating the output of a prediction at any time; and 2. budgeted batch
classification, where a fixed amount of computation is available to classify a
set of examples that can be spent unevenly across "easier" and "harder" inputs.
In contrast to most prior work, such as the popular Viola and Jones algorithm,
our approach is based on convolutional neural networks. We train multiple
classifiers with varying resource demands, which we adaptively apply during
test time. To maximally re-use computation between the classifiers, we
incorporate them as early-exits into a single deep convolutional neural network
and inter-connect them with dense connectivity. To facilitate high quality
classification early on, we use a two-dimensional multi-scale network
architecture that maintains coarse and fine level features all-throughout the
network. Experiments on three image-classification tasks demonstrate that our
framework substantially improves the existing state-of-the-art in both
settings. | http://arxiv.org/pdf/1703.09844 | Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, Kilian Q. Weinberger | cs.LG | null | null | cs.LG | 20170329 | 20180607 | [
{
"id": "1702.07780"
},
{
"id": "1702.07811"
},
{
"id": "1703.04140"
},
{
"id": "1603.08983"
},
{
"id": "1612.02297"
},
{
"id": "1604.07316"
},
{
"id": "1704.04861"
},
{
"id": "1610.02357"
}
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.