image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
((b))Empirical probability over 50 runs to converge to an optimal solution as a function ofm𝑚mitalic_mandk𝑘kitalic_k, i.e., whereℒ𝐲(𝐲105)≤10−14subscriptℒ𝐲subscript𝐲superscript105superscript1014\mathcal{L}_{\mathbf{y}}(\mathbf{y}_{10^{5}})\leq 10^{-14}caligraphic_L start_POSTSUBSCRIPT bold_y end_POSTSUBSCRIPT ( bold_y start_POSTSUBSCRIPT 10 start_POSTSUPERSCRIPT 5 end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ) ≤ 10 start_POSTSUPERSCRIPT - 14 end_POSTSUPERSCRIPT. | 2403.05395v1 | 3 |
|
Figure 2:Probability over 50 runs for a network to converge to an optimal solution for variousn𝑛nitalic_nandγ𝛾\gammaitalic_γ. | 2403.05395v1 | 4 |
|
((a))Reconstruction of an image when one uses Gaussian blur with no noise and with low level of noise. | 2403.05395v1 | 5 |
|
((a))Reconstruction of an image when one uses Gaussian blur with no noise and with low level of noise. | 2403.05395v1 | 6 |
|
((b))Evolution of the reconstruction through training with a well-conditioned operator and high level of noise. | 2403.05395v1 | 7 |
|
Figure 4:Evolution of the reconstruction error with a well-conditioned operator and high amount of noise. | 2403.05395v1 | 8 |
|
Figure 1:Comparison between DFM and our proposed GCM on the Hpatches dataset.Our approach produces more correspondences in the regions with low texture or repetitive patterns. | 2403.05388v1 | 1 |
|
Figure 4:Our proposed patch descriptor distillation strategy. We employ the R2D2 backbone as the teacher descriptor to generate a more lightweight student descriptor. BN refers to batch normalization. | 2403.05388v1 | 4 |
|
Figure 1:The value of the policy learned by FQI as a function of the size of batch dataset. The results are averaged over90909090independently collected datasets. The figures on the left, middle and right are generated using batch data that contain only1111,5555, and30303030successful trajectories respectively. The standard error of the mean is reported via the shaded region. | 2403.05385v5 | 1 |
|
Figure 2:The portion of the time that a policy learned by FQI was able to balance the pendulum for3000300030003000steps.
Results are averaged over90909090independently collected datasets and each learned policy is tested on1000100010001000initializations. The standard error of the mean is shaded. | 2403.05385v5 | 2 |
|
Figure 3:Learning curves on Asterix and Seaquest. The results are averaged over5555datasets. The shaded regions represent one standard error of the mean. One epoch contains 100k updates. | 2403.05385v5 | 3 |
|
Figure 4:Learning curves on Asterix and Seaquest. The result are averaged over 5 datasets with one standard error. One epoch contains 100k updates. | 2403.05385v5 | 4 |
|
Figure 1:Performance (mAP) of the proposed detector with DINOv2 features on the SIMD dataset, compared to YOLOv5 for different amounts of available examples per class. Robust visual features largely outperform state-of-the-art supervised methods when annotated data is limited. | 2403.05381v1 | 1 |
|
Figure 2:Building a class reference prototype for the aircraft categorypropellerwith four examples. The frozen pre-trained backbone is used to extract image representations. Then, patches overlapping box annotations are averaged into one single vector. Lastly, all four embeddings are combined into a reference vector via averaging and normalization. | 2403.05381v1 | 2 |
|
Figure 3:General diagram of our detector. An input image is fed to the RPN to generate region proposals, as well as to the backbone to extract high-level representations. Then, cosine similarity maps are generated using the features and pre-computed prototypes. For each region proposal, the mean average similarity with each prototype is computed, and the proposal is then classified as the most similar prototype class. Lastly, we discard boxes classified as a background prototype and apply non-maximum suppression. | 2403.05381v1 | 3 |
|
Figure 4:Illustrative qualitative results obtained by the proposed detector. Images on the top row correspond to the SIMD dataset, while images on the bottom belong to the DIOR dataset. | 2403.05381v1 | 4 |
|
Figure 3:The schematic of dataset creation pipelines:𝒟1subscript𝒟1\mathcal{D}_{1}caligraphic_D start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTderived from VocalSet (a),𝒟2subscript𝒟2\mathcal{D}_{2}caligraphic_D start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTand𝒟3subscript𝒟3\mathcal{D}_{3}caligraphic_D start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPTderived from Musdb18 (b), and the test dataset (𝒟4subscript𝒟4\mathcal{D}_{4}caligraphic_D start_POSTSUBSCRIPT 4 end_POSTSUBSCRIPT) creation process (c). | 2403.05380v1 | 3 |
|
Figure 4:Song-level performance curves of the models across varying thresholds, measured in terms of accuracy (a), precision (b), and recall (c). | 2403.05380v1 | 4 |
|
Figure 1:Overview of our data-efficient MIL model for AML genetic subtypes classification. We pre-train an MIL encoderfθsubscript𝑓𝜃f_{\theta}italic_f start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPTwith one of the state-of-the-art SSL models, i.e., SimCLR, SwAV, and DINO. Then we embed the trained encoder in the attention MIL architecture. For both SSL and MIL the same training dataset is used. | 2403.05379v2 | 1 |
|
Figure 2:The confusion matrix presents the test fold results of the first run for each pre-training method. While different SSL pre-training methods lead to varying class-wise performance, overall SSL pre-training performs correspondingly to fully supervised pre-training | 2403.05379v2 | 2 |
|
Figure 3:Similar mean ROC curves for MIL model resulting from different pre-training methods. Differences in the AUC score for genetic subtype prediction is within the margin of error. | 2403.05379v2 | 3 |
|
Figure 4:Averaged attention values for all labeled instances per model. Most of cell types with higher attention are malignant, while most lower attention cell types a healthy. SimCLR and DINO pretrained models provide more robust attention scores for malignant cells compared to the supervised pre-training. | 2403.05379v2 | 4 |
|
Figure 5:UMAP embedding visualization of encoder features with expert cytologist-assigned cell labels. SSL pre-trained encoder could distinguish cell-types without any label similar to fully supervised pre-trained encoder. | 2403.05379v2 | 5 |
|
Figure 1:The proposed parameter estimation pipeline.
First, we translate the agent-based model into a probabilistic generative agent-based model.
Then, we apply variational inference to get an approximate posterior of the target parameters within a given dataset. | 2403.05358v1 | 1 |
|
Figure 4:Specific errors ofεF+subscriptsuperscript𝜀𝐹\varepsilon^{+}_{F}italic_ε start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_F end_POSTSUBSCRIPT,εL+subscriptsuperscript𝜀𝐿\varepsilon^{+}_{L}italic_ε start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_L end_POSTSUBSCRIPT,εF−subscriptsuperscript𝜀𝐹\varepsilon^{-}_{F}italic_ε start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_F end_POSTSUBSCRIPT,εL−subscriptsuperscript𝜀𝐿\varepsilon^{-}_{L}italic_ε start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_L end_POSTSUBSCRIPT, and𝐫𝐫\mathbf{r}bold_ras functions of𝑵𝑵\boldsymbol{N}bold_italic_N(left),𝑻𝑻\boldsymbol{T}bold_italic_T(center), and proportion of leaders (right). The error bars represent the standard errors. | 2403.05358v1 | 4 |
|
Figure 5:Estimation time as a function of𝑵𝑵\boldsymbol{N}bold_italic_N(left) and𝑻𝑻\boldsymbol{T}bold_italic_T(right). The error bars represent standard errors. | 2403.05358v1 | 5 |
|
Figure 6:Root-mean-square error (RMSE) ofε𝜀\varepsilonitalic_εagainst some features of the ABMs trajectories, (i) number of positive interactions of followers, (ii) number of positive interactions of leaders, (iii) absolute difference betweenεF+subscriptsuperscript𝜀𝐹\varepsilon^{+}_{F}italic_ε start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_F end_POSTSUBSCRIPTandεL+subscriptsuperscript𝜀𝐿\varepsilon^{+}_{L}italic_ε start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_L end_POSTSUBSCRIPT, (iv) absolute difference betweenεF−subscriptsuperscript𝜀𝐹\varepsilon^{-}_{F}italic_ε start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_F end_POSTSUBSCRIPTandεL−subscriptsuperscript𝜀𝐿\varepsilon^{-}_{L}italic_ε start_POSTSUPERSCRIPT - end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_L end_POSTSUBSCRIPT, (v) variance of the opinion at the end of the simulation.
The regression lines are in logarithmic scale. | 2403.05358v1 | 6 |
|
(a)Better scored by the SOTA metrics (e.g., FID, KID and FDDINO-V2DINO-V2{}_{\text{DINO-V2}}start_FLOATSUBSCRIPT DINO-V2 end_FLOATSUBSCRIPT) | 2403.05352v3 | 1 |
|
Figure 2:Plausibility evaluation using Fréchet Denoised Distance.Blue area and starsvisualizing the distribution and samples of real data in the image space and in the DAE-encoded latent space;Orange area and starsillustrating the distribution and samples of generated data in the image space and in the DAE-encoded latent space. | 2403.05352v3 | 3 |
|
Figure 3:Examples of manipulated images for sensitivity test. We choose the intensity of the disturbances so that images with structural errors (i.e., mask and swap) are notably less plausible than ones with visual artifacts (i.e., salt & pepper noise and Gaussian noise). | 2403.05352v3 | 4 |
|
Figure 6:Metric comparison with increasing disturbances. The y-axis label represents the measured distance of each metric. The TD, as the most competitive metric to our FDD, performs however unstably with increasing disturbances. | 2403.05352v3 | 13 |
|
Figure 7:Metrics comparison in the task of model ranking. We color-code thediagonal linesafter the measured Pearson correlation coefficient between metric results and human judgments,dark redrefers to a strong positive correlation between metric distances and human judgments. | 2403.05352v3 | 14 |
|
Figure 1:The proposed face recognition system incorporates federated learning. Through the implementation of a secure aggregator, we empower a collective of inherently untrusting devices to collaborate and calculate an aggregate value without disclosing their individual private data. | 2403.05344v1 | 1 |
|
Figure 2:Histograms depicting the Equal Error Rate (EER) across 1000 devices are presented for the comparison between individual and federated models in the supervised systems. Notably, this evaluation focuses on models that do not utilize a secure aggregator (SA). | 2403.05344v1 | 3 |
|
Figure 3:Histograms illustrating the Equal Error Rate (EER) distribution across 1000 devices are provided for a comparison between individual and federated models in the supervised system. This analysis specifically considers models that incorporate a secure aggregator (SA). | 2403.05344v1 | 4 |
|
Figure 4:The box plot depicts the distribution of Equal Error Rates (EER) for both supervised individual and federated models across 1000 devices. The analysis considers scenarios both with and without using a Secure Aggregator (SA). Additionally, the influence of impostor selections, with and without the incorporation of Generative Adversarial Network (GAN), is highlighted. | 2403.05344v1 | 5 |
|
Figure 5:The Equal Error Rate (EER) across 1000 devices is reported for the comparison between the individual and federated models in the unsupervised system. | 2403.05344v1 | 6 |
|
Figure 6:A box plot is presented to illustrate the distribution of Equal Error Rates (EER) for unsupervised individual and federated models across 1000 devices. The analysis encompasses scenarios both with and without the utilization of a Secure Aggregator (SA). | 2403.05344v1 | 7 |
|
Figure 1:Sample output from the proposed experimental setup to evaluate the difference in performance bounds between each voting methodology. The red line in the radar chart is representative of the performance bounds (images correctly matched) and the axis represent the total number of query images in the data set. The various dimensions are useful to interpret the difference in performance of a voting scheme, in comparison to the others, such that the closer the red line is to boundary the better the performance of the voting scheme. | 2405.02297v1 | 1 |
|
Figure 2:A standard VPR ensemble set up employing several VPR methods simultaneously, that produces the top best matches by each method which are subjected to various voting schemes to observe difference in results. | 2405.02297v1 | 2 |
|
Figure 3:Difference in performance bounds of each voting methodology including Plurality, Condorcet, Contingent, Broda Count and Instant Run Off voting, in terms of query images correctly matched for different data sets : 17Places (top left), Livingroom (top right), Corridor (center left), CrossSeasons (center right), ESSEX3IN1 (bottom left) and GardensPoint (bottom right). | 2405.02297v1 | 3 |
|
Figure 4:Performance in terms of PR curves for each voting methodology including Plurality, Condorcet, Contingent Voting, Broda Count and Instant Run Off voting, in terms of query images correctly matched for different data sets : 17Places (top left), Livingroom (top right), Corridor (center left), CrossSeasons (center right), ESSEX3IN1 (bottom left) and GardensPoint (bottom right). | 2405.02297v1 | 4 |
|
Figure 5:Pairwise comparisons between the voting methods have been considered. A sign convention is used to present the results: a positive value of Z indicates that the first method of the pair outperforms the second one, whereas a negative Z score has the opposite meaning. Corresponding to the z-scores the legend represents the confidence intervals starting with the highest confidence intervals being green to lowest being red. | 2405.02297v1 | 5 |
|
Figure 1:Trade-off between the number of Giga Multiply-Accumulate (GMAC) Operations and Jaccard prediction score of our architecture, U-Net[14]and state-of-the-art ELU-Net[15] | 2403.05340v1 | 1 |
|
Figure 2:The novelty of our architecture lies in its ability to generate high resolution predictions with very compressed inputs. | 2403.05340v1 | 2 |
|
Figure 3:Overview of our architecture.
Our architecture extends the conventional ”U” shape of the network to produce higher-resolution outputs. In training, we compare the output at different sizes to the correspondingly reshaped ground truths. | 2403.05340v1 | 3 |
|
Figure 4:Detailed illustration of our architecture up-scaling layers with skip connections for the case of a16×16161616\times 1616 × 16input image with one class. | 2403.05340v1 | 4 |
|
Figure 5:A comparison between input resolution and prediction quality in Jaccard of the different networks on Decathlon. | 2403.05340v1 | 5 |
|
Figure 7:Example results of our architecture and ELU-Net for comparison. The a) column shows the T2 brain MRIs in their original resolution. The b) column shows the ground truth in the original resolution and the c) and d) columns the T2 brain MRIs and ground truths in input resolution respectively, the resolution is specified in the top left corner in column d), and the e) column shows the prediction of ELU-Net), and the f) column shows the prediction of our architecture. | 2403.05340v1 | 7 |
|
Figure 8:Example results of our architecture and ELU-Net for comparison. The a) column shows the T2 brain MRIs in their original resolution. The b) column shows the ground truth in the original resolution and the c) and d) columns the T2 brain MRIs and ground truths in input resolution respectively, the resolution is specified in the top left corner in column d), and the e) column shows the prediction of ELU-Net), and the f) column shows the prediction of our architecture. | 2403.05340v1 | 8 |
|
Figure 1:Extraction of explanatory signals from PBMs. Yellow boxes:actual task input. Blue boxes:trigger tokens. Pink box: prediction token. Orange boxes: last hidden representations of PBM. Green box: predicted label (converted by verbalizer, e.g., positive→→\rightarrow→great, negative→→\rightarrow→bad). | 2403.05338v1 | 1 |
|
Figure 3:TheF1subscript𝐹1F_{1}italic_F start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTscores of models trained with different sizes. From top to bottom: TSE and e-SNLI. | 2403.05338v1 | 7 |
|
Figure 1:Single layer update with hook layer (residual connections are omitted).∥.∥\parallel.\parallel∥ . ∥means calculate the L2-norm over the keys’ dimension (m𝑚mitalic_m). For each updating of a single batch edits, the temporary hook layer is used at the beginning to ensureΔΔ\Deltaroman_Δis computed based onWhlsuperscriptsubscript𝑊ℎ𝑙W_{h}^{l}italic_W start_POSTSUBSCRIPT italic_h end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_l end_POSTSUPERSCRIPT. After the weights update, the validated hook layer is applied to determine whether to use the original layer or hook layer for each token. This process can be implemented iteratively to support consecutive batch editing. Note that the temporary hook layer weight of a new iteration is copied from the validated hook layer weight of the previous iteration. So, the validated hook layer keeps track of the updated layer from previous edits by retaining the weight from the previous iteration. | 2403.05330v3 | 1 |
|
Figure 2:Multiple layer update with hook layer (Attention module and the first layer of FFN are omitted). The value vectorvisubscript𝑣𝑖v_{i}italic_v start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTis first computed at the last editing layer, and then we iteratively insert a fraction of the residual to each editing layer (I, II, III) using Eq.13. Since changing one layer would affect the activations of downstream layers, recollection of the activations is conducted after each iteration. At the beginning, temporary hook layers are initialized to all editing layers. Once the hook layer weight is updated, it is replaced by the validated hook layer (1, 2, 3). | 2403.05330v3 | 2 |
|
Figure 3:Difference between the z-score entry to the updated keyZkeylsubscriptsuperscript𝑍𝑙𝑘𝑒𝑦Z^{l}_{key}italic_Z start_POSTSUPERSCRIPT italic_l end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_k italic_e italic_y end_POSTSUBSCRIPTand average ofZlsuperscript𝑍𝑙Z^{l}italic_Z start_POSTSUPERSCRIPT italic_l end_POSTSUPERSCRIPT. The x-axis represents the sample index. | 2403.05330v3 | 3 |
|
Figure 10:Performance comparisons on the different number of editing layers. Layers are selected from the critical path identified inMeng et al. (2023). | 2403.05330v3 | 10 |
|
Figure 1:Visualization of our coarse-grained and fine-grained prediction results. The first row shows the ground truth and prediction for two coarse-grained samples, while the second row displays the ground truth and prediction for the same two samples at a fine-grained level. Better viewed when zoomed in. | 2403.05329v2 | 1 |
|
Figure 2:Comparison of our method with one of the existing SOTA multi-modal baseline[48]under challenging samples. The first row compares M-baseline[48]with our proposed OccFusion for the coarse occupancy prediction task, while the second row compares M-CONet[48]with our Active M-CONet. Better viewed when zoomed in. | 2403.05329v2 | 2 |
|
Figure 3:The overall architecture of our method. Raw LiDAR points are processed by a 3D encoder to extract voxelized features, which, concatenated with point coordinates, serve as queries. Multi-view image features, obtained directly through a 2D encoder from surround-view images, act as keys. Enhanced point clouds are then subjected to point-to-point fusion, resulting in multi-modal 3D voxel features. An active decoder adaptively refines predictions in challenging areas. | 2403.05329v2 | 3 |
|
Figure 4:Details of the OccFusion module. After pre-sampling, 3D reference points are projected onto images as (2D) reference points. Note that synthetic point clouds (points within circles) do not contribute to LiDAR feature generation. Due to overlapping fields of view among cameras, a single 3D reference point may correspond to multiple reference points upon projection. Features corresponding to reference points are averaged to derive a feature for each 3D reference point, which are then averaged to obtain a multi-modal feature for a voxel. | 2403.05329v2 | 4 |
|
Figure 5:Active coarse to fine pipeline. We refine features only for voxels with greater uncertainty. | 2403.05329v2 | 5 |
|
Figure 6:Active training method. In the training stage, we train the model using the training set sampled from the previous stage. In the resampling stage, we use the model trained in the training stage to score the loss on thefulltraining set, selecting the top K percent of samples to form the training set for the next training cycle. | 2403.05329v2 | 6 |
|
Figure 4:Visualization of outlier prediction onF3DosubscriptF3Do\text{F3D}_{\text{o}}F3D start_POSTSUBSCRIPT o end_POSTSUBSCRIPTdataset.Black: Accurate prediction.Red: Outliers. Top row: Outliers defined as EPE > 0.30. Bottom row: Outliers predicted by Uncertainty. | 2403.05327v3 | 4 |
|
Figure 2:Proposed pipeline. During the feature extractor fine-tuning stage (left), a pre-trained feature extractor model is fed with image patches coming from a larger context window. A random subset of the patches’ feature vector representations is masked, and a Transformer encoder with a predictor network is used to predict the masked instances’ feature vector representations produced by a frozen teacher network, minimizing anl1subscript𝑙1l_{1}italic_l start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTloss. For the downstream task training stage (right), the Transformer and the predictor networks are discarded, and the fine-tuned feature extractor can be used in any Multiple Instance Learning pipeline. | 2403.05325v1 | 4 |
|
Figure 3:Heatmap visualizations of the cosine similarity between the feature vector of a patch from a metastasis region and the rest of the feature vectors of the slide. | 2403.05325v1 | 5 |
|
Figure 1:Comparison between the performance of Peacock and mBlip models on SEED-Benchmark dimensions. | 2403.01031v2 | 1 |
|
Figure 2:Peacock InstructBLIP architecture: Integrates instruction-specific visual features using Q-Former and a frozen pretrained image encoder. | 2403.01031v2 | 2 |
|
Figure 2:Peacock InstructBLIP architecture: Integrates instruction-specific visual features using Q-Former and a frozen pretrained image encoder. | 2403.01031v2 | 3 |
|
Figure 3:Peacock LLaVA architecture: Combines a pretrained frozen vision encoder with trained Arabic LLMs via an MLP bridge. | 2403.01031v2 | 4 |
|
Figure 4:Our data filtering pipeline. After translating the data through Google Cloud API, we obtain the embeddings of both the original and translated samples using the multilingual sentence embedding model LaBSE. For each sample, we calculate the cosine similarity between the two extracted embeddings and reject samples under an 80% threshold. | 2403.01031v2 | 5 |
|
Figure 5:Examples of responses from Peacock and GPT-4V regarding an image related to Yemeni culture. | 2403.01031v2 | 6 |
|
Figure 6:Both Peacock and GPT-4V accurately respond to a question in the Egyptian dialect. While GPT-4V provides a slightly more detailed answer, it does so in MSA. In contrast, Peacock’s response is in the same Egyptian dialect as the question. | 2403.01031v2 | 7 |
|
Figure 7:Human evaluation results on the accuracy and authenticity of model responses to questions about images in Egyptian dialect. "IB-AraLLaMA " denotes ourInstructBlip-AraLLaMAmodel, and "IB-AceGPT" refers to ourInstructBlip-AceGPTmodel. "GPT-4V DI" is the GPT-4V model explicitly instructed to respond in the Egyptian dialect."GPT-4V" represents the GPT-4V model, which is given a question in the Egyptian dialect, similar to how Peacock models are instructed. | 2403.01031v2 | 8 |
|
Figure A.1:Three examples illustrating the variations in translation quality from English to Arabic, ranging from good to moderate to poor, as indicated by the similarity scores. | 2403.01031v2 | 9 |
|
Figure A.2:This collection of images showcases a curated subset selected from Henna dataset, representing 11 Arab countries, and capturing the essence of traditional food, local customs, historical monuments, everyday activities, and distinctive architecture that characterize the diverse and rich heritage of each region. | 2403.01031v2 | 10 |
|
Figure A.3:Dataset Generation Example using GPT-4V. This figure demonstrates the process of generating a question-answer dataset for an attraction in Yemen as an example. For each site, an image and its corresponding Wikipedia article were used to provide GPT-4V with rich contextual information. The model then generated ten contextually relevant questions and answers per image. | 2403.01031v2 | 11 |
|
(b)Ten questions and answers generated using the image from5(a)with Henna generation pipeline. All questions and answers have been translated into English to demonstrate the quality of the generated data. | 2403.01031v2 | 15 |
|
(a)High-scoring evaluations: The answer from InstructBlip-AraLLaMA is accurate, earning a perfect score from GPT-4. mBlip’s response is also rated as reasonable. | 2403.01031v2 | 16 |
|
(a)High-scoring evaluations: The answer from InstructBlip-AraLLaMA is accurate, earning a perfect score from GPT-4. mBlip’s response is also rated as reasonable. | 2403.01031v2 | 17 |
|
(b)Moderate-scoring evaluations: Here, InstructBlip-AraLLaMA’s response, while accurate, doesn’t match the gold standard, resulting in a moderate GPT-4 score. mBlip’s scoring reflects a proper evaluation. | 2403.01031v2 | 18 |
|
(c)Low-scoring evaluations: InstructBlip-AraLLaMA’s response is incorrect, leading to a low GPT-4 score. Conversely, mBlip is awarded a moderate score, indicating a more accurate response. | 2403.01031v2 | 19 |
|
Figure A.14:Comparative responses fromInstructBlip-AraLLaMAandInstructBlip-AceGPTmodels alongside the correct MSA answers for four selected examples, three from Henna and the last one from LLaVA-Bench, demonstrating the models’ proficiency in Egyptian dialect. | 2403.01031v2 | 27 |
|
Figure A.15:A demonstration of Peacock model’s advantage over GPT-4V in responding to queries in the Egyptian dialect, with the former responding in the dialect and the latter in MSA. A secondary example shows both models replying in MSA when queried in MSA. | 2403.01031v2 | 28 |
|
Figure 1:Time evolution of the operator⟨σ+σ−⟩delimited-⟨⟩subscript𝜎subscript𝜎\langle\sigma_{+}\sigma_{-}\rangle⟨ italic_σ start_POSTSUBSCRIPT + end_POSTSUBSCRIPT italic_σ start_POSTSUBSCRIPT - end_POSTSUBSCRIPT ⟩for a high (infrequent measurement) and low (frequent measurement) values ofgzsubscript𝑔𝑧g_{z}italic_g start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT. Depending on the measurement rate, the system exhibits a Zeno effect with frequent measurements (the orange solid line) or it undergoes oscillations when the measurement rate is low (the blue dashed line). | 2403.01024v1 | 1 |
|
Figure 2:Occupation probabilities of Fock states|nσ⟩∈|00⟩…|21⟩ket𝑛𝜎ket00…ket21|n\sigma\rangle\in|00\rangle\dots|21\rangle| italic_n italic_σ ⟩ ∈ | 00 ⟩ … | 21 ⟩as a function of cavity driving amplitudeβ𝛽\betaitalic_β. | 2403.01024v1 | 2 |
|
Figure 3:Sketch of the RC system with a measurement-controlled quantum dynamics. The pivotal component of the reservoir is a cavity-atom system with continuously monitored states. The neural activations of the reservoir are given by the Fock states of the atom-cavity quantum system. The reservoir is coherently driven using a signal defined by the discrete points of the input dataset. The classified readouts of the reservoir are processed by means of a linear regression technique. | 2403.01024v1 | 3 |
|
Figure 4:(a)Input data generated from a random array representing either a sinusoidal or square waveform. Results of the classification of the sinusoidal and square waveform by the reservoir with(b)8 and(c)16 neurons. The target is shown in solid green line and the reservoir prediction in dashed red line. | 2403.01024v1 | 4 |
|
Figure 5:The root-mean-square error (RMSE) (the blue square markers) and the accuracy (the red circular markers) obtained for the sinusoidal-square waveform classification task as a function of the number of neurons in the reservoir. | 2403.01024v1 | 5 |
|
Figure 6:Performance of the reservoir as the function of the measurement rate. The optimal RMSE is achieved at the measurement rate of 5 arb. units. | 2403.01024v1 | 6 |
|
Figure 7:Generative mode operation exemplified by a free-running forecast of MGTS. In this figure, we compare the output of the reservoir (the dashed-red line) with the target MGTS (the solid green line). The reservoir has 16 neurons and it was trained on several cycles of MGTS variations. Note that the reservoir was not presented with the ground truth MGTS data to make the forecast. The comparison with the ground truth is needed only to evaluate the accuracy of the forecast. | 2403.01024v1 | 7 |
|
Figure 9:Output of the RC system trained to predict a damped harmonic oscillator. The free-running forecast made by the the RC system is denoted by the dashed red line. The solid green line denotes the ground truth. | 2403.01024v1 | 9 |
|
(b)A sample Gaussian process embedding ofS1superscript𝑆1S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTintoℝ2superscriptℝ2\mathbb{R}^{2}blackboard_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT | 2403.07929v1 | 3 |
|
(a)Embeddings ofS1×3.5S1superscript𝑆13.5superscript𝑆1S^{1}\times 3.5S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT × 3.5 italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT | 2403.07929v1 | 4 |
|
(a)Embeddings ofS1×3.5S1superscript𝑆13.5superscript𝑆1S^{1}\times 3.5S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT × 3.5 italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT | 2403.07929v1 | 5 |
|
(b)Sample embedding ofS1superscript𝑆1S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTplus two outliers with DMS | 2403.07929v1 | 9 |
|
(c)Sample embedding ofS1superscript𝑆1S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTplus two outliers with DMS, outliers removed | 2403.07929v1 | 10 |
|
Figure 4:Three sample embeddings ofS1superscript𝑆1S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTplus two outliers with GPS | 2403.07929v1 | 11 |
|
(b)Results for point clouds with all 4 types of Gaussian process embeddings, comparison with Euclidean distance | 2403.07929v1 | 17 |
|
(a)Embeddings ofS1×3.5S1superscript𝑆13.5superscript𝑆1S^{1}\times 3.5S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT × 3.5 italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTintoℝ8superscriptℝ8\mathbb{R}^{8}blackboard_R start_POSTSUPERSCRIPT 8 end_POSTSUPERSCRIPT,p𝑝pitalic_pvarying,L𝐿Litalic_Lmeasured with Euclidean distance | 2403.07929v1 | 18 |
|
(a)Embeddings ofS1×3.5S1superscript𝑆13.5superscript𝑆1S^{1}\times 3.5S^{1}italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT × 3.5 italic_S start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTintoℝ8superscriptℝ8\mathbb{R}^{8}blackboard_R start_POSTSUPERSCRIPT 8 end_POSTSUPERSCRIPT,p𝑝pitalic_pvarying,L𝐿Litalic_Lmeasured with Euclidean distance | 2403.07929v1 | 19 |
Subsets and Splits