id,figure_id,figure_type,paper_title,authors,arxiv_id,category,year,license,citation,original_caption,small_caption,medium_caption,large_caption,extracted_text,text_score,clip_score,composite_score,image_width,image_height,aspect_ratio,image_hash,image,pdf_link
arxiv_0000001,Figure 1,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 1: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-BN+SH and the
softmax baseline trained on retrieval task for different model dimensions and sequence lengths. The reduction in
computation and memory when using our models improves with sequence length. When scaling up the model,
our methods remain significantly more beneficial than the baseline. Time Series experiment, on this LRA benchmark, Attention-BN and Attention-SH both outperform
the softmax attention on most five tasks. Moreover, Attention-BN+SH, which combines these two
attention mechanisms, results in the most accurate models on average across tasks. Specifically, for
the retrieval task, the most challenging task with the largest sequence length in the LRA benchmark,
Attention-BN+SH achieve a remarkable improvement of more than 1.5% over the baseline. Image Classification on Imagenet. We corroborate the advantage of our proposed attention over
the baseline softmax attention when scaled up for the large-scale ImageNet image classification
task. We summarize the results in Table 3. The Deit model (Touvron et al., 2021) equiped with the
Attention-BN yields better performance than the softmax baseline. Meanwhile, Attention-SH/BN+SH
Deit perform on par with the baseline while being more efficient. These results, together with other
results above justify the benefits of our proposed methods across various tasks and data modalities,
proving the effectiveness of our primal-dual approach to develop new attentions. 4
EMPIRICAL ANALYSIS Efficiency Analysis. The Attention-BN+SH not only improves the accuracy of the model remarkably
but also help reduce the computational and memory cost significantly. Fig.1 presents the efficiency
benefits of our Attention-BN+SH trained on the retrieval task when the model dimension D and
sequence lengths N grow. The efficiency advantage of our model increase as N increase. In addition,
the scaled-up models (with large D) remains significantly more efficient than the baseline. When
the model dimension is 64 and sequence length is 4096, which is the standard configuration of the
task, the model’s FLOPS, in both training and inference, reduce almost 25%, whereas the reductions
for memory usage in training and testing are 31.9% and 47.3%, respectively. Notably, this efficient
model also outperforms the baseline with more than 1.5% improvement in accuracy. These results
prove the benefits of applying the Attention-BN+SH for long-sequence tasks and large-scale models. New Attentions Helps Reduce Head Redundancy. We compute the average L2 distances between
heads to analyze the attention diversity. Given our trained models for the retrieval task, the layer-
average mean and standard deviation of distances between heads are reported in Table 4. All our
introduced attentions attain greater L2 distances compared to the baseline, reducing the risk of
learning redundant heads. In particular, Attention-SH has the highest head difference, indicating the
model’s attention patterns are most spread out between heads. 8",A diagram of a series of different.,A detailed layout showing the differences between sequence length and distance of a training course or testimore.,"A comprehensive technical explanation of the difference between a training and sequence length of each exercise, using the same method as described by the figure below, and the data in the table below, 1 / 2 / 3 / 6 / 4.","Figure 1: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-BN+SH and the
softmax baseline trained on retrieval task for different model dimensions and sequence lengths. The reduction in
computation and memory when using our models improves with sequence length. When scaling up the model,
our methods remain significantly more beneficial than the baseline. Time Series experiment, on this LRA benchmark, Attention-BN and Attention-SH both outperform
the softmax attention on most five tasks. Moreover, Attention-BN+SH, which combines these two
attention mechanisms, results in the most accurate models on average across tasks. Specifically, for
the retrieval task, the most challenging task with the largest sequence length in the LRA benchmark,
Attention-BN+SH achieve a remarkable improvement of more than 1.5% over the baseline. Image Classification on Imagenet. We corroborate the advantage of our proposed attention over
the baseline softmax attention when scaled up for the large-scale ImageNet image classification
task. We summarize the results in Table 3. The Deit model (Touvron et al., 2021) equiped with the
Attention-BN yields better performance than the softmax baseline. Meanwhile, Attention-SH/BN+SH
Deit perform on par with the baseline while being more efficient. These results, together with other
results above justify the benefits of our proposed methods across various tasks and data modalities,
proving the effectiveness of our primal-dual approach to develop new attentions. 4
EMPIRICAL ANALYSIS Efficiency Analysis. The Attention-BN+SH not only improves the accuracy of the model remarkably
but also help reduce the computational and memory cost significantly. Fig.1 presents the efficiency
benefits of our Attention-BN+SH trained on the retrieval task when the model dimension D and
sequence lengths N grow. The efficiency advantage of our model increase as N increase. In addition,
the scaled-up models (with large D) remains significantly more efficient than the baseline. When
the model dimension is 64 and sequence length is 4096, which is the standard configuration of the
task, the model’s FLOPS, in both training and inference, reduce almost 25%, whereas the reductions
for memory usage in training and testing are 31.9% and 47.3%, respectively. Notably, this efficient
model also outperforms the baseline with more than 1.5% improvement in accuracy. These results
prove the benefits of applying the Attention-BN+SH for long-sequence tasks and large-scale models. New Attentions Helps Reduce Head Redundancy. We compute the average L2 distances between
heads to analyze the attention diversity. Given our trained models for the retrieval task, the layer-
average mean and standard deviation of distances between heads are reported in Table 4. All our
introduced attentions attain greater L2 distances compared to the baseline, reducing the risk of
learning redundant heads. In particular, Attention-SH has the highest head difference, indicating the
model’s attention patterns are most spread out between heads. 8",0.81,0.3181,0.5641,1920,960,2.0,74ca7c78a79ad2c717fd16b17f6ce8de,images/2024/arxiv_0000001.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000002,Figure 2,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 1: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-BN+SH and the
softmax baseline trained on retrieval task for different model dimensions and sequence lengths. The reduction in
computation and memory when using our models improves with sequence length. When scaling up the model,
our methods remain significantly more beneficial than the baseline.",A diagram of the number of different.,A detailed layout showing the different stages of the cell phone ' s battery lifecycle from 1900 to 2000.,"A comprehensive technical explanation of the different types of the cell phone ' s battery lifecycle from the first to the second generation of cellular phones, including the new and older models, 1970 - 2012 and present in the first half.","Figure 1: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-BN+SH and the
softmax baseline trained on retrieval task for different model dimensions and sequence lengths. The reduction in
computation and memory when using our models improves with sequence length. When scaling up the model,
our methods remain significantly more beneficial than the baseline.",0.725,0.2113,0.4681,714,201,3.552,8edc6d8b507203cde72be8aca4cee9db,images/2024/arxiv_0000002.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000003,Figure 3,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 2: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-SH and the softmax
attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.",A diagram of the stages in testing and.,"A detailed layout showing the differences between the sequence length and the training time for a dog, and the distance of the training.","A comprehensive technical explanation of the two basic training curves for the b - testing and sequence lengths of each one source image from the department of the university of applied materials of engineering, 2013 / eocnepto /.","Figure 2: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-SH and the softmax
attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.",0.81,0.3023,0.5562,1920,960,2.0,dd7dd292bbe0e10ea6db12d7aaf40bea,images/2024/arxiv_0000003.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000004,Figure 4,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 3: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-BN+SH and
the Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence
lengths. When using our models, the reduction in computation and memory improves with sequence length.
When scaling up the model with greater model dimension, our methods remain significantly more efficient than
the baseline.",A diagram of the different stages of a.,"A detailed layout showing the differences between sequence length and training time in the training system, with a single line graph.","A comprehensive technical explanation of the sequence length and training of a dog ' s teeth and teeth, with their corresponding corresponding lengths and the same length as shown below diagram in the following data graph above the chart below each.","Figure 3: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-BN+SH and
the Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence
lengths. When using our models, the reduction in computation and memory improves with sequence length.
When scaling up the model with greater model dimension, our methods remain significantly more efficient than
the baseline.",0.87,0.3135,0.5917,1920,960,2.0,e9b8091a6b5ce7d36a569ad03198035a,images/2024/arxiv_0000004.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000005,Figure 5,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 4: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-SH and the
Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.
When using our models, the reduction in computation and memory improves with sequence length. When
scaling up the model with greater model dimension, our methods remain significantly more efficient than the
baseline.
Attention-SH. Fig.2 shows the efficiency benefits of our Attention-SH when trained on the retrieval
task. Same as in the case of Attention-SH+BN, the efficiency benefits of our Attention-SH over the
baseline Softmax attention grows when N and D increase. Sparse Attention-SH/BN+SH. Fig.3 and Fig.4 show that the efficiency advantages of our Sparse
Attention-BN+SH and Sparse Attention-SH, respectively, increase as the model dimension D and
sequence length N grow. All models are trained on the LRA retrieval task. In addition to the efficiency
advantage, the Sparse Attention-BN+SH also significantly outperforms the Sparse Attention baseline
in terms of accuracy in this task (79.86% vs. 78.20%) while the Sparse Attention-SH achieves a
comparable result to the baseline. More accuracy advantages of the Sparse Attention-BN/SH/BN+SH
over the Sparse Attention baseline are given in Table 10. D
DERIVING SOFTMAX ATTENTION. Choosing the appropriate h(x) and Φ(x) allows us to derive the popular softmax attention given in
Eqn. 1 and 2. In particular, if we choose h(x) := PN
j Φ(x)T Φ(kj), Eqn. 10 becomes PN
j=1 Φ(x)⊤Φ(kj)vj
PN
j′ Φ(x)T Φ(kj′)
+ b.
(27) N
X Φ(x)⊤Φ(kj)
PN
j′ Φ(x)T Φ(kj′)
vj + b = f(x) = j=1",A diagram of the different stages of a.,A detailed layout showing the differences between a training and sequence length of the results in a single file.,"A comprehensive technical explanation of the sequence length and distance of a training course for a specific course, and a specific training course, as described by the following the data file file from the data source to the data editor ' s.","Figure 4: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-SH and the
Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.
When using our models, the reduction in computation and memory improves with sequence length. When
scaling up the model with greater model dimension, our methods remain significantly more efficient than the
baseline.
Attention-SH. Fig.2 shows the efficiency benefits of our Attention-SH when trained on the retrieval
task. Same as in the case of Attention-SH+BN, the efficiency benefits of our Attention-SH over the
baseline Softmax attention grows when N and D increase. Sparse Attention-SH/BN+SH. Fig.3 and Fig.4 show that the efficiency advantages of our Sparse
Attention-BN+SH and Sparse Attention-SH, respectively, increase as the model dimension D and
sequence length N grow. All models are trained on the LRA retrieval task. In addition to the efficiency
advantage, the Sparse Attention-BN+SH also significantly outperforms the Sparse Attention baseline
in terms of accuracy in this task (79.86% vs. 78.20%) while the Sparse Attention-SH achieves a
comparable result to the baseline. More accuracy advantages of the Sparse Attention-BN/SH/BN+SH
over the Sparse Attention baseline are given in Table 10. D
DERIVING SOFTMAX ATTENTION. Choosing the appropriate h(x) and Φ(x) allows us to derive the popular softmax attention given in
Eqn. 1 and 2. In particular, if we choose h(x) := PN
j Φ(x)T Φ(kj), Eqn. 10 becomes PN
j=1 Φ(x)⊤Φ(kj)vj
PN
j′ Φ(x)T Φ(kj′)
+ b.
(27) N
X Φ(x)⊤Φ(kj)
PN
j′ Φ(x)T Φ(kj′)
vj + b = f(x) = j=1",0.87,0.319,0.5945,1920,960,2.0,30ed925affa7bab4353c8b3a900f5be4,images/2024/arxiv_0000005.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000006,Figure 6,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 3: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-BN+SH and
the Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence
lengths. When using our models, the reduction in computation and memory improves with sequence length.
When scaling up the model with greater model dimension, our methods remain significantly more efficient than
the baseline.",A diagram of the differences between.,A detailed layout showing the differences between the sequence length and the training time for each of the two teams.,"A comprehensive technical explanation of the sequence length and technique used in this training course, including the following steps for each step, with the following one line graphed up at the same time of the following stages at the teste.","Figure 3: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-BN+SH and
the Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence
lengths. When using our models, the reduction in computation and memory improves with sequence length.
When scaling up the model with greater model dimension, our methods remain significantly more efficient than
the baseline.",0.81,0.2885,0.5493,1920,960,2.0,7d412f7765289b832b4c272efe6e722e,images/2024/arxiv_0000006.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000007,Figure 7,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 4: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-SH and the
Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.
When using our models, the reduction in computation and memory improves with sequence length. When
scaling up the model with greater model dimension, our methods remain significantly more efficient than the
baseline.
Attention-SH. Fig.2 shows the efficiency benefits of our Attention-SH when trained on the retrieval
task. Same as in the case of Attention-SH+BN, the efficiency benefits of our Attention-SH over the
baseline Softmax attention grows when N and D increase. Sparse Attention-SH/BN+SH. Fig.3 and Fig.4 show that the efficiency advantages of our Sparse
Attention-BN+SH and Sparse Attention-SH, respectively, increase as the model dimension D and
sequence length N grow. All models are trained on the LRA retrieval task. In addition to the efficiency
advantage, the Sparse Attention-BN+SH also significantly outperforms the Sparse Attention baseline
in terms of accuracy in this task (79.86% vs. 78.20%) while the Sparse Attention-SH achieves a
comparable result to the baseline. More accuracy advantages of the Sparse Attention-BN/SH/BN+SH
over the Sparse Attention baseline are given in Table 10. D
DERIVING SOFTMAX ATTENTION. Choosing the appropriate h(x) and Φ(x) allows us to derive the popular softmax attention given in
Eqn. 1 and 2. In particular, if we choose h(x) := PN
j Φ(x)T Φ(kj), Eqn. 10 becomes PN
j=1 Φ(x)⊤Φ(kj)vj
PN
j′ Φ(x)T Φ(kj′)
+ b.
(27) N
X Φ(x)⊤Φ(kj)
PN
j′ Φ(x)T Φ(kj′)
vj + b = f(x) = j=1",A diagram of the sequence length and.,"A detailed layout showing the difference of a training and sequence length for each course, and the same number of training.","A comprehensive technical explanation of the sequence length and training of a dog using a computer screen, and a graphed line graph on a white background with data from the same time and the same height as shown in each line,.","Figure 4: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-SH and the
Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.
When using our models, the reduction in computation and memory improves with sequence length. When
scaling up the model with greater model dimension, our methods remain significantly more efficient than the
baseline.
Attention-SH. Fig.2 shows the efficiency benefits of our Attention-SH when trained on the retrieval
task. Same as in the case of Attention-SH+BN, the efficiency benefits of our Attention-SH over the
baseline Softmax attention grows when N and D increase. Sparse Attention-SH/BN+SH. Fig.3 and Fig.4 show that the efficiency advantages of our Sparse
Attention-BN+SH and Sparse Attention-SH, respectively, increase as the model dimension D and
sequence length N grow. All models are trained on the LRA retrieval task. In addition to the efficiency
advantage, the Sparse Attention-BN+SH also significantly outperforms the Sparse Attention baseline
in terms of accuracy in this task (79.86% vs. 78.20%) while the Sparse Attention-SH achieves a
comparable result to the baseline. More accuracy advantages of the Sparse Attention-BN/SH/BN+SH
over the Sparse Attention baseline are given in Table 10. D
DERIVING SOFTMAX ATTENTION. Choosing the appropriate h(x) and Φ(x) allows us to derive the popular softmax attention given in
Eqn. 1 and 2. In particular, if we choose h(x) := PN
j Φ(x)T Φ(kj), Eqn. 10 becomes PN
j=1 Φ(x)⊤Φ(kj)vj
PN
j′ Φ(x)T Φ(kj′)
+ b.
(27) N
X Φ(x)⊤Φ(kj)
PN
j′ Φ(x)T Φ(kj′)
vj + b = f(x) = j=1",0.81,0.2759,0.543,1920,960,2.0,ffb613e48d3e31fd0453e2efcf94c8cc,images/2024/arxiv_0000007.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000008,Figure 8,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 2: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-SH and the softmax
attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.",A diagram of different types of the.,A detailed layout showing the different levels of the light intensity of the leds and their corresponding brightness.,"A comprehensive technical explanation of the different types of the data processing system for each type of machine, from 1 to 8 years old and up to 7 years old, in the same stages of development, with the same time,.","Figure 2: (Left) FLOPS ratios and (Right) memory usage ratios between the Attention-SH and the softmax
attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.",0.75,0.2171,0.4835,754,212,3.557,478ea0df619f7cc792cd3d761ef4e25e,images/2024/arxiv_0000008.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000009,Figure 9,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 3: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-BN+SH and
the Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence
lengths. When using our models, the reduction in computation and memory improves with sequence length.
When scaling up the model with greater model dimension, our methods remain significantly more efficient than
the baseline.",A diagram of the different stages of.,A detailed layout showing the number of training steps for each type of trainer at different stages of training.,"A comprehensive technical explanation of the different levels of training and training for each individual person in a sports team, including a running coach, a squaturic, a training, a squashing, and a - kayaking,.","Figure 3: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-BN+SH and
the Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence
lengths. When using our models, the reduction in computation and memory improves with sequence length.
When scaling up the model with greater model dimension, our methods remain significantly more efficient than
the baseline.",0.81,0.221,0.5155,754,200,3.77,37bc31f53271d3253df5cd92bd7f870b,images/2024/arxiv_0000009.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000010,Figure 10,scientific_figure,A Primal-Dual Framework for Transformers and Neural Networks,arXiv Research Authors,2406.13781v1,cs,2024,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2024). A Primal-Dual Framework for Transformers and Neural Networks. arXiv:2406.13781v1,"Figure 4: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-SH and the
Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.
When using our models, the reduction in computation and memory improves with sequence length. When
scaling up the model with greater model dimension, our methods remain significantly more efficient than the
baseline.
Attention-SH. Fig.2 shows the efficiency benefits of our Attention-SH when trained on the retrieval
task. Same as in the case of Attention-SH+BN, the efficiency benefits of our Attention-SH over the
baseline Softmax attention grows when N and D increase.",A diagram of the stages of a cell.,"A detailed layout showing the different stages of a proteining process, including the different levels of protein.","A comprehensive technical explanation of the different levels of heat in the world ' s most radioactive - like substances for the first time, and how much is that it actually changed??? - - - based on - page - page.","Figure 4: (Left) FLOPS ratios and (Right) memory usage ratios between the Sparse Attention-SH and the
Sparse Attention baseline trained on the LRA retrieval task for different model dimensions and sequence lengths.
When using our models, the reduction in computation and memory improves with sequence length. When
scaling up the model with greater model dimension, our methods remain significantly more efficient than the
baseline.
Attention-SH. Fig.2 shows the efficiency benefits of our Attention-SH when trained on the retrieval
task. Same as in the case of Attention-SH+BN, the efficiency benefits of our Attention-SH over the
baseline Softmax attention grows when N and D increase.",0.75,0.2323,0.4911,754,204,3.696,581e4eef55dcfa64e765158cfacc090e,images/2024/arxiv_0000010.png,https://arxiv.org/pdf/2406.13781v1.pdf
arxiv_0000011,Figure 11,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,Figures and captions,A diagram of the process for a camera.,"A detailed layout showing the various components of a laser machine and its equipment, including the main parts.","A comprehensive technical explanation of a micro - laser system for the use of a microscope and other instruments, including a small camera, a microscope, and a computer and a large screen monitor and a microscope are shown in the same.",Figures and captions,0.75,0.2861,0.5181,1349,1461,0.923,3b4f4e6cc81e93c3d0010f11f68b04d5,images/2025/arxiv_0000011.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000012,Figure 12,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 1. (a) Schematic layout: Overview of the snapshot multi-spectral imaging system, 
illustrating the arrangement of the circular LED array, Digital Micromirror Device (DMD), 
optical components (collimator, diffuser, lenses), and the monochrome imaging camera. (b) 
Experimental set-up.",A diagram of a camera and a computer.,"A detailed layout showing the components of a robotic arm and equipment for the project, including a camera.","A comprehensive technical explanation of a large scale laser machine, with its components labeled and labelled in english and english, for the use of engineers and technicianss to study the area of their needs on the market and customers.","Figure 1. (a) Schematic layout: Overview of the snapshot multi-spectral imaging system, 
illustrating the arrangement of the circular LED array, Digital Micromirror Device (DMD), 
optical components (collimator, diffuser, lenses), and the monochrome imaging camera. (b) 
Experimental set-up.",0.7357,0.3655,0.5506,1224,1155,1.06,4fef3b79b0d4553799783caf67825617,images/2025/arxiv_0000012.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000013,Figure 13,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 2. (a-b) Visualizations of image reconstruction results of two test image patterns 
illuminated under different wavelengths. Each column refers to an image with the monochrome 
defocused sensor input, network output and multi-spectral target (ground truth). Pseudo colors 
represent the corresponding illumination LED color (center wavelength).",A diagram of various colors of.,"A detailed layout showing different color schemes of the same pattern in various colors, including red, green, blue, and yellow.","A comprehensive technical explanation of the different color combinations of the multi - colored samples in the image below, from the first to the next two years of this time of the year, and then the same period, the same time.","Figure 2. (a-b) Visualizations of image reconstruction results of two test image patterns 
illuminated under different wavelengths. Each column refers to an image with the monochrome 
defocused sensor input, network output and multi-spectral target (ground truth). Pseudo colors 
represent the corresponding illumination LED color (center wavelength).",0.75,0.2312,0.4906,1371,1106,1.24,eaed630072830ba41f5d086ac64eaf9a,images/2025/arxiv_0000013.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000014,Figure 14,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 2. (a-b) Visualizations of image reconstruction results of two test image patterns 
illuminated under different wavelengths. Each column refers to an image with the monochrome 
defocused sensor input, network output and multi-spectral target (ground truth). Pseudo colors 
represent the corresponding illumination LED color (center wavelength).",A diagram of the different color.,"A detailed layout showing the different color variations of each cell phone, in different colors and sizes,.","A comprehensive technical explanation of the multicolored structure of the cell membranes of the human and verteopia of the cellular cell membrane - based on this sample, weltronomic model is shown below the image.","Figure 2. (a-b) Visualizations of image reconstruction results of two test image patterns 
illuminated under different wavelengths. Each column refers to an image with the monochrome 
defocused sensor input, network output and multi-spectral target (ground truth). Pseudo colors 
represent the corresponding illumination LED color (center wavelength).",0.7357,0.242,0.4889,1018,823,1.237,bbfe75c8b96da20c04731d6389948e97,images/2025/arxiv_0000014.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000015,Figure 15,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 3. Visualization of image reconstruction results under various illumination 
configurations. Each column displays the predicted image alongside the corresponding target 
(ground truth) image for different combinations of activated LEDs that were on.",A diagram of the different types of.,"A detailed layout showing the different properties of the structure of the cell phone, including the signal and the signal.","A comprehensive technical explanation of the different types of light in a photograph of a multicolored background, with the following colors and the same number of light that is visible in the image, and red, black, red,.","Figure 3. Visualization of image reconstruction results under various illumination 
configurations. Each column displays the predicted image alongside the corresponding target 
(ground truth) image for different combinations of activated LEDs that were on.",0.75,0.269,0.5095,1440,941,1.53,88c02e349f7ae008c22f13324ce7dbaf,images/2025/arxiv_0000015.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000016,Figure 16,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 3. Visualization of image reconstruction results under various illumination 
configurations. Each column displays the predicted image alongside the corresponding target 
(ground truth) image for different combinations of activated LEDs that were on.",A diagram of the different colors of a.,A detailed layout showing the different effects of different colors in the image and how they are used to create them.,"A comprehensive technical explanation of the different effects of color in photoshopped images, from a series of images with varying color schemes, to a series with the same image and texture and size, each color scheme, each.","Figure 3. Visualization of image reconstruction results under various illumination 
configurations. Each column displays the predicted image alongside the corresponding target 
(ground truth) image for different combinations of activated LEDs that were on.",0.75,0.3012,0.5256,937,620,1.511,aca53fa72a09f79a2c151051556dc5c4,images/2025/arxiv_0000016.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000017,Figure 17,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 4. Image reconstruction performance analysis. (a) Confusion matrix of spectral channel 
accuracy. (b) Classification metrics against the number of concurrent LED illuminations: Plots of 
sensitivity, specificity, and F1 score against the number of concurrent LED illuminations. (c) 
Energy difference values as a function of the number of concurrent LED illuminations; boxplots 
depict the distribution of energy differences for various numbers of concurrent illuminations, 
indicating the shift in the prediction accuracy and variance. (d) Energy difference values as a 
function of the illumination wavelength. We used light green to signify “losses” and light blue 
for “leakages”.",A diagram of the different types of.,A detailed layout showing the effects of the different types of energy in the environment and how it affects.,"A comprehensive technical explanation of the different effects of different types of energy in a system that is not very efficient, and the results are very attractive to us $ 3, $ 5, 000, 000 / $ 1, 000.","Figure 4. Image reconstruction performance analysis. (a) Confusion matrix of spectral channel 
accuracy. (b) Classification metrics against the number of concurrent LED illuminations: Plots of 
sensitivity, specificity, and F1 score against the number of concurrent LED illuminations. (c) 
Energy difference values as a function of the number of concurrent LED illuminations; boxplots 
depict the distribution of energy differences for various numbers of concurrent illuminations, 
indicating the shift in the prediction accuracy and variance. (d) Energy difference values as a 
function of the illumination wavelength. We used light green to signify “losses” and light blue 
for “leakages”.",0.7393,0.2537,0.4965,1440,1235,1.166,359fed75205db9382f5268b5bfd69771,images/2025/arxiv_0000017.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000018,Figure 18,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 4. Image reconstruction performance analysis. (a) Confusion matrix of spectral channel 
accuracy. (b) Classification metrics against the number of concurrent LED illuminations: Plots of 
sensitivity, specificity, and F1 score against the number of concurrent LED illuminations. (c) 
Energy difference values as a function of the number of concurrent LED illuminations; boxplots 
depict the distribution of energy differences for various numbers of concurrent illuminations, 
indicating the shift in the prediction accuracy and variance. (d) Energy difference values as a 
function of the illumination wavelength. We used light green to signify “losses” and light blue 
for “leakages”.",A diagram of the different types of.,A detailed layout showing the different types of energy that are present in the same region of the earth.,"A comprehensive technical explanation of the energy and energy - generating process in the environment of a city, including energy - efficient buildings, solar - powered buildings, and wind - driven - driven systems, and - driven power - off -.","Figure 4. Image reconstruction performance analysis. (a) Confusion matrix of spectral channel 
accuracy. (b) Classification metrics against the number of concurrent LED illuminations: Plots of 
sensitivity, specificity, and F1 score against the number of concurrent LED illuminations. (c) 
Energy difference values as a function of the number of concurrent LED illuminations; boxplots 
depict the distribution of energy differences for various numbers of concurrent illuminations, 
indicating the shift in the prediction accuracy and variance. (d) Energy difference values as a 
function of the illumination wavelength. We used light green to signify “losses” and light blue 
for “leakages”.",0.725,0.2157,0.4703,974,833,1.169,692a32725c8ac88e9714e74741cfeb61,images/2025/arxiv_0000018.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000019,Figure 19,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 5. Distributions of (a) PSNR, (b) RMSE, (c) SSIM and (d) NMSE metrics for the 
reconstructed images at different illumination wavelengths.",A diagram of the distribution of water.,"A detailed layout showing the mean of the weight of each individual figure, from different data points to a number of different measurements.","A comprehensive technical explanation of the hypothe - and - weight distribution for different types of water, including the following three phases of the cycle of the process and the process of the hydro cycle, the process,.","Figure 5. Distributions of (a) PSNR, (b) RMSE, (c) SSIM and (d) NMSE metrics for the 
reconstructed images at different illumination wavelengths.",0.81,0.2684,0.5392,1371,789,1.738,8c075c515dee190e59857e8224a38bd4,images/2025/arxiv_0000019.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000020,Figure 20,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 5. Distributions of (a) PSNR, (b) RMSE, (c) SSIM and (d) NMSE metrics for the 
reconstructed images at different illumination wavelengths.",A diagram of the different phases of a.,A detailed layout showing different types of the scatters and patterns for different areas in the region.,"A comprehensive technical explanation of the glycolized datasete using the rx - 3 and rxp - 2 algorithms and rp - 4 - 6f - 6 - 5 - biting functions representing all details, specifications, and configurations of the system components in full.","Figure 5. Distributions of (a) PSNR, (b) RMSE, (c) SSIM and (d) NMSE metrics for the 
reconstructed images at different illumination wavelengths.",0.725,0.231,0.478,957,550,1.74,4d39a18ad8be903db5d87b7c24ed1f27,images/2025/arxiv_0000020.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000021,Figure 21,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 6. (a) Distributions of structural quality metrics for the reconstructed multi-spectral 
images with respect to the number of concurrent LED illuminations. (b) Confusion matrices with 
different numbers of concurrent LED illuminations.",A diagram of the number of people who.,"A detailed layout showing the number of individual and individual patients in each hospital room, including a number of different.","A comprehensive technical explanation of the different types of water in the world ' s oceans and oceans, including the number of lakes, land, and land surface areas, and water, from sea level to sea level, and area.","Figure 6. (a) Distributions of structural quality metrics for the reconstructed multi-spectral 
images with respect to the number of concurrent LED illuminations. (b) Confusion matrices with 
different numbers of concurrent LED illuminations.",0.75,0.2607,0.5053,1371,1465,0.936,4e072d790924a09b959dc272d431b47d,images/2025/arxiv_0000021.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000022,Figure 22,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 6. (a) Distributions of structural quality metrics for the reconstructed multi-spectral 
images with respect to the number of concurrent LED illuminations. (b) Confusion matrices with 
different numbers of concurrent LED illuminations.",A diagram of the number and type of.,"A detailed layout showing the number of different types of cell phone phones, with the same number of cellular phones.","A comprehensive technical explanation of the results of the four different stages of the experiment in the following diagrams, with the following results of each one in the previous model image and the following steps on the next slide below slides.","Figure 6. (a) Distributions of structural quality metrics for the reconstructed multi-spectral 
images with respect to the number of concurrent LED illuminations. (b) Confusion matrices with 
different numbers of concurrent LED illuminations.",0.75,0.2405,0.4952,957,992,0.965,92f63691f0cb4daeb06259f7123ef76c,images/2025/arxiv_0000022.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000023,Figure 23,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 7. Network architecture. (a) Multi-spectral Fourier Imager Network (mFIN). (b) Dense 
links: each output tensor of the dSPAF group is appended and fed to the subsequent one. (c) 
Detailed schematic of dSPAF modules. See the Methods section for details.",A diagram of the process of the.,A detailed layout showing the process of the application for a remote control system with different functions and processes.,"A comprehensive technical explanation of the application of an aspx / gdsm - based system for a remote access network in a remote area of a remote control system, with remote access and remote control systems, from a remote.","Figure 7. Network architecture. (a) Multi-spectral Fourier Imager Network (mFIN). (b) Dense 
links: each output tensor of the dSPAF group is appended and fed to the subsequent one. (c) 
Detailed schematic of dSPAF modules. See the Methods section for details.",0.75,0.2499,0.5,1440,667,2.159,f9b32c7898a536895feffca4d41f82eb,images/2025/arxiv_0000023.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000024,Figure 24,scientific_figure,Snapshot multi-spectral imaging through defocusing and a Fourier imager network,arXiv Research Authors,2501.14287v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Snapshot multi-spectral imaging through defocusing and a Fourier imager network. arXiv:2501.14287v1,"Figure 7. Network architecture. (a) Multi-spectral Fourier Imager Network (mFIN). (b) Dense 
links: each output tensor of the dSPAF group is appended and fed to the subsequent one. (c) 
Detailed schematic of dSPAF modules. See the Methods section for details.",A diagram of the process of a computer.,"A detailed layout showing the use of the sdr - s and the application for the sdrs that shows key features, attributes, and data points in detail for.","A comprehensive technical explanation of the application of the p2p - 5 module for the application to the new p2psp - 4 module, based on - 5p - 3p - 6p - p2 - 8p representing all details, specifications, and configurations of the system components.","Figure 7. Network architecture. (a) Multi-spectral Fourier Imager Network (mFIN). (b) Dense 
links: each output tensor of the dSPAF group is appended and fed to the subsequent one. (c) 
Detailed schematic of dSPAF modules. See the Methods section for details.",0.81,0.2693,0.5396,957,449,2.131,f38d365511ca98015aeeb9745912912d,images/2025/arxiv_0000024.png,https://arxiv.org/pdf/2501.14287v1.pdf
arxiv_0000025,Figure 25,scientific_figure,Fast Solvers for Discrete Diffusion Models: Theory and Applications of High-Order Algorithms,arXiv Research Authors,2502.00234v2,cs,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Fast Solvers for Discrete Diffusion Models: Theory and Applications of High-Order Algorithms. arXiv:2502.00234v2,"Figure 1: Left: Application of the uniformization algorithm to discrete diffusion models for text
generation. The x-axis denotes the time of the backward process, and the y-axis denotes the
frequency of jumps (NFE). Perplexity convergence occurs before the NFE grows unbounded. Right:
Comparison between τ-leaping and the proposed second-order schemes (θ-RK-2 and θ-Trapezoidal).",A diagram of a computer screen showing.,A detailed layout showing the differences between the three different types of the model and the same type of model.,"A comprehensive technical explanation of the different types of neural processing techniques for a computer application, including the following data and the following results of the process in the same process - time of the image, from the.","Figure 1: Left: Application of the uniformization algorithm to discrete diffusion models for text
generation. The x-axis denotes the time of the backward process, and the y-axis denotes the
frequency of jumps (NFE). Perplexity convergence occurs before the NFE grows unbounded. Right:
Comparison between τ-leaping and the proposed second-order schemes (θ-RK-2 and θ-Trapezoidal).",0.81,0.202,0.506,782,262,2.985,ffb833c014022b869127c3255c9e4423,images/2025/arxiv_0000025.png,https://arxiv.org/pdf/2502.00234v2.pdf
arxiv_0000026,Figure 26,scientific_figure,Fast Solvers for Discrete Diffusion Models: Theory and Applications of High-Order Algorithms,arXiv Research Authors,2502.00234v2,cs,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Fast Solvers for Discrete Diffusion Models: Theory and Applications of High-Order Algorithms. arXiv:2502.00234v2,"Figure 2: Empirical KL divergence between the true
and generated distribution of the toy model vs. number
of steps. Data are fitted with linear regression with 95%
confidence interval by bootstrapping.",A diagram of the number of steps.,"A detailed layout showing the number of steps to be taken in a step - by - step process that shows key features, attributes, and data points in.","A comprehensive technical explanation of a step curve in the computer system, including the steps and the number of steps that will be taken to the computer machine for each step in the next step, and the process of the next time.","Figure 2: Empirical KL divergence between the true
and generated distribution of the toy model vs. number
of steps. Data are fitted with linear regression with 95%
confidence interval by bootstrapping.",0.81,0.2789,0.5444,455,249,1.827,32569438aca37b8e7e692411d9885df9,images/2025/arxiv_0000026.png,https://arxiv.org/pdf/2502.00234v2.pdf
arxiv_0000027,Figure 27,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 1: Reinforcement learning framework for adaptive control of the over-relaxation
parameter in LBGK. Each agent samples an action from its policy, with actions in-
terpolated across the grid and applied to the environment. Agents receive local state
observations and a global reward based on the alignment of the coarse-grained simula-
tion energy spectrum with the target DNS spectrum.",A diagram of the process of a software.,A detailed layout showing the components of the system in which a sample action is generated to be completed.,"A comprehensive technical explanation of the model for the experiment in which the data is stored and displayed in a flow diagram, as well as examples of the corresponding data in the image, the sequences and other elements, the actual.","Figure 1: Reinforcement learning framework for adaptive control of the over-relaxation
parameter in LBGK. Each agent samples an action from its policy, with actions in-
terpolated across the grid and applied to the environment. Agents receive local state
observations and a global reward based on the alignment of the coarse-grained simula-
tion energy spectrum with the target DNS spectrum.",0.7393,0.2737,0.5065,1352,1126,1.201,88e80358066f9e7ecd0a4e0fc62e330a,images/2025/arxiv_0000027.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000028,Figure 28,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 1: Reinforcement learning framework for adaptive control of the over-relaxation
parameter in LBGK. Each agent samples an action from its policy, with actions in-
terpolated across the grid and applied to the environment. Agents receive local state
observations and a global reward based on the alignment of the coarse-grained simula-
tion energy spectrum with the target DNS spectrum.",A diagram of a group of different.,"A detailed layout showing the flow of the process of an experiment, with the sequence and its corresponding components.","A comprehensive technical explanation of the algorithm for a single - action system, including the same function as the other two - step sequencer, is shown in figure - based on a computer simulations and a computer graphicsized image.","Figure 1: Reinforcement learning framework for adaptive control of the over-relaxation
parameter in LBGK. Each agent samples an action from its policy, with actions in-
terpolated across the grid and applied to the environment. Agents receive local state
observations and a global reward based on the alignment of the coarse-grained simula-
tion energy spectrum with the target DNS spectrum.",0.75,0.264,0.507,1191,688,1.731,20156cdfe50f74af15c722a2c9160219,images/2025/arxiv_0000028.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000029,Figure 29,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 2: Example for a policy network parametrization in an environment with local
and cooperating agents. Figure 2a shows the network architecture for a fully local agent.
The agent only receives the state at its location which in this case is a six dimensional
vector. The network then consists of two fully convolutional layers and has two output
neurons, namely the mean µ and standard deviation σ. These parametrize a normal
distribution from which the agent can sample actions πi(a|s) ∼N(µ, σ). Figure 2b show
the vectorized version of Fig. 2a. The experience of all agents is combined and handled
as a single evaluation of a fully convolutional network.",A diagram of the three sections of an.,A detailed layout showing the two different sections of a building with two different angles and one side of it.,"A comprehensive technical explanation of the optical system for scanning and viewing the image of a rectangled object from different angles in perspective and perspective, to which the image shows the image is not visible and the same image from.","Figure 2: Example for a policy network parametrization in an environment with local
and cooperating agents. Figure 2a shows the network architecture for a fully local agent.
The agent only receives the state at its location which in this case is a six dimensional
vector. The network then consists of two fully convolutional layers and has two output
neurons, namely the mean µ and standard deviation σ. These parametrize a normal
distribution from which the agent can sample actions πi(a|s) ∼N(µ, σ). Figure 2b show
the vectorized version of Fig. 2a. The experience of all agents is combined and handled
as a single evaluation of a fully convolutional network.",0.75,0.2258,0.4879,893,219,4.078,132df63f8ab556450296baa400e9242a,images/2025/arxiv_0000029.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000030,Figure 30,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 3: Neural network architecture for a centralized critic network used in the learning
phase by the MARL PPO algorithm. The Network receives a state S as input, which
is compressed by six convolutional and three fully connected layers, and outputs a value
function estimate V (S).",A diagram of a plane with a number of.,"A detailed layout showing the various parts of a wooden structure and a line of wooden beams, with different sections.","A comprehensive technical explanation of the 3d image of a box and its contents and components in a single - dimensional view of it, which is an object that appears to be a box or something that is a box, that.","Figure 3: Neural network architecture for a centralized critic network used in the learning
phase by the MARL PPO algorithm. The Network receives a state S as input, which
is compressed by six convolutional and three fully connected layers, and outputs a value
function estimate V (S).",0.75,0.2718,0.5109,777,236,3.292,5cac0cae195ecf684f44c8e79a4ec861,images/2025/arxiv_0000030.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000031,Figure 31,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 4: Evaluation of trained models on Kolmogorov flow at Re = 104. (4a) show the
vorticity correlation of models with the DNS. All three models are able to stabilize the
simulation. (4b) shows the energy spectra scaled by k5, averaged over the second half
of the simulation T ∈[113, 227]. All three models reproduce the target spectrum of the
DNS, with small deviations at higher wave numbers.",A diagram of a graph showing the.,A detailed layout showing the results of different time - lapsing and multi - dimensional datas in the data.,"A comprehensive technical explanation of the graphing process in the data visual system for the data modeling software, with the following data visual tool - based image and the data processing tool kit included in the following versions of the.","Figure 4: Evaluation of trained models on Kolmogorov flow at Re = 104. (4a) show the
vorticity correlation of models with the DNS. All three models are able to stabilize the
simulation. (4b) shows the energy spectra scaled by k5, averaged over the second half
of the simulation T ∈[113, 227]. All three models reproduce the target spectrum of the
DNS, with small deviations at higher wave numbers.",0.7957,0.2742,0.535,2539,1575,1.612,fa7fc1eb7f7a6bab19773bbe3bc123ee,images/2025/arxiv_0000031.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000032,Figure 32,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",A diagram of the average flow of water.,A detailed layout showing the distribution of the water vapors in the atmosphere and the temperature of the atmosphere.,"A comprehensive technical explanation of the water cycle for the seawater in the uk and europe, from a data sheet for the waterworkser ' s perspective, 2012 - 2015 - 2016 - 06, with a - 20 - 07 representing all details, specifications, and.","Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",0.81,0.2484,0.5292,2577,1576,1.635,bab79cc710f4f2deee1fafdb58737d08,images/2025/arxiv_0000032.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000033,Figure 33,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",A diagram of the number of people who.,A detailed layout showing the distribution of the number of different types of gas in the world ' s atmosphere.,"A comprehensive technical explanation of the time series for the first half of 2009 - 2013, with a line graph of the same time and the last half of 2013 - 2013 - 2012 - 2016 - 2019 years, and the year.","Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",0.8064,0.2675,0.537,2539,1575,1.612,bbb50b7d8968580ed8ea5f5e3cc5c0ce,images/2025/arxiv_0000033.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000034,Figure 34,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",A diagram of the temperature curve of.,"A detailed layout showing the average number of people in a group, and how they are using the data that shows key features, attributes, and data.","A comprehensive technical explanation of the mean on the watermark model for the current waveforms, with the following data of the time and volume of the stream - based data at the same time, as well as described.","Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",0.81,0.2517,0.5309,2577,1576,1.635,1c7fcba14479a01047fc82f2acb0df2f,images/2025/arxiv_0000034.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000035,Figure 35,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 4: Evaluation of trained models on Kolmogorov flow at Re = 104. (4a) show the
vorticity correlation of models with the DNS. All three models are able to stabilize the
simulation. (4b) shows the energy spectra scaled by k5, averaged over the second half
of the simulation T ∈[113, 227]. All three models reproduce the target spectrum of the
DNS, with small deviations at higher wave numbers.",A diagram of the same line and the.,"A detailed layout showing the various stages of a wave and its characteristics in different directions, including the wave.","A comprehensive technical explanation of the data visual for a single - phase system of data visual in a single file system, with data visual of the same stage 3 / 4 / 3 / 3, 5 / 6 / 8 / 6 representing all details, specifications, and.","Figure 4: Evaluation of trained models on Kolmogorov flow at Re = 104. (4a) show the
vorticity correlation of models with the DNS. All three models are able to stabilize the
simulation. (4b) shows the energy spectra scaled by k5, averaged over the second half
of the simulation T ∈[113, 227]. All three models reproduce the target spectrum of the
DNS, with small deviations at higher wave numbers.",0.75,0.2251,0.4875,1191,458,2.6,2656148faef68afe6eb2561b0d287ea3,images/2025/arxiv_0000035.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000036,Figure 36,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",A diagram of a curve and a line graph.,A detailed layout showing the differences between the two data fields on this model and the same data field.,"A comprehensive technical explanation of the data visual data for the two different systems, including the data and the data processing process and the information processing process on the same one line graphing process is also the same line with.","Figure 5: Evaluation of trained models on an unforced decaying flow at Re = 104.(5a)
show the vorticity correlation of models with the DNS. (5b) shows the energy spectra
scaled by k5, averaged over the second half of the simulation T ∈[113, 227].",0.7957,0.2438,0.5197,1191,365,3.263,fa1162b0549c469d7bb23fd5135cd812,images/2025/arxiv_0000036.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000037,Figure 37,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 6: Evaluation of trained models on Kolmogorov flow at Re = 105. (6a) show the
vorticity correlation of models with the DNS. (6b) shows the energy spectra scaled by k5,
averaged over the second half of the simulation T ∈[113, 227].",A diagram of the flow of water from a.,A detailed layout showing the various types of water vapors and their characteristics in the atmosphere of the earth.,"A comprehensive technical explanation of the temperature curve and its relationship to the temperature of the water source, including the temperature, temperature, and temperature, in the water cycle, and heat area of the earth ' s surface ' s.","Figure 6: Evaluation of trained models on Kolmogorov flow at Re = 105. (6a) show the
vorticity correlation of models with the DNS. (6b) shows the energy spectra scaled by k5,
averaged over the second half of the simulation T ∈[113, 227].",0.75,0.2178,0.4839,2577,1576,1.635,a93964e0c253144a245d0134ead3e551,images/2025/arxiv_0000037.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000038,Figure 38,scientific_figure,Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning,arXiv Research Authors,2504.14422v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Optimal Lattice Boltzmann Closures through Multi-Agent Reinforcement Learning. arXiv:2504.14422v1,"Figure 7: Comparison of the vorticity fields of all models on the three test cases: Kol-
mogorov flow at Re = 104 (7a), decaying unforced flow at Re = 104 (7b), and Kolmogorov
flow at Re = 105 (7c).",A diagram of the image shows a variety.,"A detailed layout showing the formation of a large group of cells that are in a single celled cell that shows key features, attributes, and data.","A comprehensive technical explanation of the structure and function of a cell membranes, including its surface and the number of cells in the cells, as well as described by the following the image, the image is shown.","Figure 7: Comparison of the vorticity fields of all models on the three test cases: Kol-
mogorov flow at Re = 104 (7a), decaying unforced flow at Re = 104 (7b), and Kolmogorov
flow at Re = 105 (7c).",0.81,0.2562,0.5331,1416,2030,0.698,7d128eb2d68e2803e6ad22840f50abac,images/2025/arxiv_0000038.png,https://arxiv.org/pdf/2504.14422v1.pdf
arxiv_0000039,Figure 39,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,"FIG. 1. Illustration of the three force-generation strategies for a sin-
gle unimodal distribution (e.g., a bond length): Bottom panel: Three
points sampled from the data distribution. When atomistic forces
are available, they can be used directly in subsequent applications.
Middle panel: Each of the three samples is perturbed by Gaussian
noise (noise kernel), broadening the distribution. The original sam-
ple points, which now serve as the centers of the Gaussian kernels,
are displayed semi-transparent. Forces computed from this noising
process correspond to the distorted distribution and thus inherit its
local inaccuracies. Top panel: We apply a learned reverse noise ker-
nel (via a conditional flow) to correct the perturbed samples, yielding
a distribution closer to the original. Forces derived from this reverse
kernel result in fewer local distortion in subsequent applications.",A diagram of the different types of.,"A detailed layout showing the motion of a wave in different stages, including a transverse noise kernel and a noise kernel.","A comprehensive technical explanation of the noise and noise in the sound waves of the sound source from the sound wave, the noise is the result of the amplitude of the waves in the noise moving waves of sound waves on the sound.","FIG. 1. Illustration of the three force-generation strategies for a sin-
gle unimodal distribution (e.g., a bond length): Bottom panel: Three
points sampled from the data distribution. When atomistic forces
are available, they can be used directly in subsequent applications.
Middle panel: Each of the three samples is perturbed by Gaussian
noise (noise kernel), broadening the distribution. The original sam-
ple points, which now serve as the centers of the Gaussian kernels,
are displayed semi-transparent. Forces computed from this noising
process correspond to the distorted distribution and thus inherit its
local inaccuracies. Top panel: We apply a learned reverse noise ker-
nel (via a conditional flow) to correct the perturbed samples, yielding
a distribution closer to the original. Forces derived from this reverse
kernel result in fewer local distortion in subsequent applications.",0.75,0.3689,0.5595,479,748,0.64,9a4fa0ed63f3576d5cdde8b7196dd300,images/2025/arxiv_0000039.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000040,Figure 40,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,FIG. 2. Ramachandran plots for alanine dipeptide for the different models trained on 10% of the training set.,A diagram of a green and blue.,"A detailed layout showing the structure of a molecule with four arrows and two ends, and two red and one blue.","A comprehensive technical explanation of the structure of a hydroic compound, 3d - generated image courtesy university of technology and applied graphics, university of engineering, university at albany, 2013 - new hampshire, cambridge, london.",FIG. 2. Ramachandran plots for alanine dipeptide for the different models trained on 10% of the training set.,0.7429,0.3208,0.5318,1280,1072,1.194,b130da4aa72872133c488df0e4fe391d,images/2025/arxiv_0000040.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000041,Figure 41,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,"FIG. 3. Comparison of global-feature accuracy (dihedral-angle distributions or TICA projections) versus local-feature accuracy (bond-length
distributions). Left column: The three benchmark systems and their CG representations: Alanine dipeptide, Chignolin, and Trp-cage. Each
row corresponds to one system, with panels showing results for the different training set sizes. Pareto-optimal models are indicated by bold
markers, with membership determined solely based on the mean performance. The Atomistic model, having been trained with atomistic forces,
is excluded from the Pareto sets and is presented solely as a reference. Furthermore, we report the atomistic results only on the full training set,
since its performance on the reduced datasets is inferior to the other models. Error bars represent the standard deviation across three CGSchNet
models; for the kernel-based methods, each CGSchNet model was trained on a different set of forces generated with the respective kernels.
Note that the error bars for the bond Wasserstein distances are too small to be visible.",A diagram of a 3d model of an animal.,A detailed layout showing the structure of a cell phone with multiple colored balls and needles attached to it.,"A comprehensive technical explanation of the 3d model of an anti - cell system, with multiple structures and a multicolored structure, including the nucleus - platelets and the same colored spheres on the outer surface as well as.","FIG. 3. Comparison of global-feature accuracy (dihedral-angle distributions or TICA projections) versus local-feature accuracy (bond-length
distributions). Left column: The three benchmark systems and their CG representations: Alanine dipeptide, Chignolin, and Trp-cage. Each
row corresponds to one system, with panels showing results for the different training set sizes. Pareto-optimal models are indicated by bold
markers, with membership determined solely based on the mean performance. The Atomistic model, having been trained with atomistic forces,
is excluded from the Pareto sets and is presented solely as a reference. Furthermore, we report the atomistic results only on the full training set,
since its performance on the reduced datasets is inferior to the other models. Error bars represent the standard deviation across three CGSchNet
models; for the kernel-based methods, each CGSchNet model was trained on a different set of forces generated with the respective kernels.
Note that the error bars for the bond Wasserstein distances are too small to be visible.",0.7464,0.2562,0.5013,1280,960,1.333,a8a13557afffd30340745c31a358eae9,images/2025/arxiv_0000041.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000042,Figure 42,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,FIG. 2. Ramachandran plots for alanine dipeptide for the different models trained on 10% of the training set.,A diagram of the structure of a green.,"A detailed layout showing the structure of a multicolored liquid molecule, with multiple small metal structures.","A comprehensive technical explanation of the mechanism of a multi - dimensional molecule, based on the structure of a protein cell, and its structural properties, including the structure, and functions, and function, and processes, together,.",FIG. 2. Ramachandran plots for alanine dipeptide for the different models trained on 10% of the training set.,0.75,0.3204,0.5352,1280,1072,1.194,36fad7a42455d9535b85f0dad993a0a6,images/2025/arxiv_0000042.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000043,Figure 43,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,FIG. 2. Ramachandran plots for alanine dipeptide for the different models trained on 10% of the training set.,A diagram of the temperature and.,"A detailed layout showing the various types of waves in different directions, and their corresponding colors and sizes.","A comprehensive technical explanation of the high - resolution image of a waveformer ' s speed and time of the waveformers in a single - stream, with different waves and several different stages of motion, each waveforms.",FIG. 2. Ramachandran plots for alanine dipeptide for the different models trained on 10% of the training set.,0.75,0.2398,0.4949,854,184,4.641,ff2bb5569e0b423ea74feb63b807506d,images/2025/arxiv_0000043.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000044,Figure 44,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,"FIG. 3. Comparison of global-feature accuracy (dihedral-angle distributions or TICA projections) versus local-feature accuracy (bond-length
distributions). Left column: The three benchmark systems and their CG representations: Alanine dipeptide, Chignolin, and Trp-cage. Each
row corresponds to one system, with panels showing results for the different training set sizes. Pareto-optimal models are indicated by bold
markers, with membership determined solely based on the mean performance. The Atomistic model, having been trained with atomistic forces,
is excluded from the Pareto sets and is presented solely as a reference. Furthermore, we report the atomistic results only on the full training set,
since its performance on the reduced datasets is inferior to the other models. Error bars represent the standard deviation across three CGSchNet
models; for the kernel-based methods, each CGSchNet model was trained on a different set of forces generated with the respective kernels.
Note that the error bars for the bond Wasserstein distances are too small to be visible.",A diagram of the different types of.,"A detailed layout showing the different types of data for different areas of the body and their characteristics, including.","A comprehensive technical explanation of the different types of the data stored in the data storage area of a computer system, including data storage and data storage for all data storages and storages, including the datas, as well.","FIG. 3. Comparison of global-feature accuracy (dihedral-angle distributions or TICA projections) versus local-feature accuracy (bond-length
distributions). Left column: The three benchmark systems and their CG representations: Alanine dipeptide, Chignolin, and Trp-cage. Each
row corresponds to one system, with panels showing results for the different training set sizes. Pareto-optimal models are indicated by bold
markers, with membership determined solely based on the mean performance. The Atomistic model, having been trained with atomistic forces,
is excluded from the Pareto sets and is presented solely as a reference. Furthermore, we report the atomistic results only on the full training set,
since its performance on the reduced datasets is inferior to the other models. Error bars represent the standard deviation across three CGSchNet
models; for the kernel-based methods, each CGSchNet model was trained on a different set of forces generated with the respective kernels.
Note that the error bars for the bond Wasserstein distances are too small to be visible.",0.75,0.2752,0.5126,760,598,1.271,8734dec9bd05d90c6e687bae66457df9,images/2025/arxiv_0000044.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000045,Figure 45,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,FIG. 4. TICA projections for Chignolin for the different models trained on 10% of the training set.,A diagram of the different types of.,A detailed layout showing the distribution of the ionation of the plasmas in the human body and the number of the photon.,"A comprehensive technical explanation of the time - laps for multiple images of a cell phone showing the different stages of cell phone life cycle and the same time of the cell phone - laps, as well as seen from the time.",FIG. 4. TICA projections for Chignolin for the different models trained on 10% of the training set.,0.81,0.2413,0.5257,789,193,4.088,043df06d26db1afba78d491f8b9c5fe0,images/2025/arxiv_0000045.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000046,Figure 46,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,"FIG. 5. Top row: Distribution of minimized energies for backmapped CG configurations across the three benchmark systems, comparing all
models. Bottom row: Corresponding example bond distributions.",A diagram of different types of energy.,A detailed layout showing the various types of the different energy levels of an ion in the atmosphere and the corresponding.,"A comprehensive technical explanation of the different types of chiromin and chiromine in the various phases of the cycle of chironin, chiroquin, chirominoid and chironon energy, and chirodonol representing all details, specifications, and.","FIG. 5. Top row: Distribution of minimized energies for backmapped CG configurations across the three benchmark systems, comparing all
models. Bottom row: Corresponding example bond distributions.",0.75,0.2698,0.5099,956,464,2.06,8f17371ebf4a404dab91193439546ad5,images/2025/arxiv_0000046.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000047,Figure 47,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,"FIG. 6. Ramachandran plots for the AV and KS dipeptides, samples
generated with different models. The pretrained Timewarp model is
used to obtain forces via the transition kernel.",A diagram of the different types of.,"A detailed layout showing the differences between the two phases of the h2h2 and h3h5 that shows key features, attributes, and data points in detail.","A comprehensive technical explanation of the time - laps of an experiment using the tim - t2s and the timeframes t4s - t5s t7s t3s t1s t9s representing all details, specifications, and configurations of the system components in full context with.","FIG. 6. Ramachandran plots for the AV and KS dipeptides, samples
generated with different models. The pretrained Timewarp model is
used to obtain forces via the transition kernel.",0.81,0.2005,0.5052,495,272,1.82,baaf205aed0047988640b147fc6c7fbc,images/2025/arxiv_0000047.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000048,Figure 48,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,"FIG. 7. Ramachandran plots for alanine dipeptide at varying training-set sizes: top row 100% of the data; middle row 10%; bottom row 2%.
Note that all Atomistic model simulations diverged at the 2% data level.",A diagram of the different phases of a.,A detailed layout showing the different types of heat and temperature for the area that is located on the surface.,"A comprehensive technical explanation of the time series for the tms - 1 model and the tm - 3 model, including the time, time, and current timeframes, and timeframe, and temperatures, the time representing all details, specifications, and.","FIG. 7. Ramachandran plots for alanine dipeptide at varying training-set sizes: top row 100% of the data; middle row 10%; bottom row 2%.
Note that all Atomistic model simulations diverged at the 2% data level.",0.75,0.2435,0.4968,956,448,2.134,1a90ebc70d48a136f3ef63fe7fb97824,images/2025/arxiv_0000048.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000049,Figure 49,scientific_figure,Operator Forces For Coarse-Grained Molecular Dynamics,arXiv Research Authors,2506.19628v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Operator Forces For Coarse-Grained Molecular Dynamics. arXiv:2506.19628v1,FIG. 8. TICA plots for Chignolin at varying training-set sizes: top row 100% of the data; middle row 10%; bottom row 2%.,A diagram of the various types of.,A detailed layout showing the different stages of the human brain and how it is used to study the different areas.,"A comprehensive technical explanation of the results of a multi - dimensional pattern of human eye and nose tissues, from the study of the study for the study in the study on - to the study atc, jl representing all details, specifications, and.",FIG. 8. TICA plots for Chignolin at varying training-set sizes: top row 100% of the data; middle row 10%; bottom row 2%.,0.75,0.1812,0.4656,956,412,2.32,3b17b3e39201ea16c17d760ffb6cb885,images/2025/arxiv_0000049.png,https://arxiv.org/pdf/2506.19628v1.pdf
arxiv_0000050,Figure 50,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 1. Representative results of our method (PnP-CM) across a diverse set of inverse problems. Left to right: linear inverse problems
(Gaussian deblurring, super-resolution, inpainting) and nonlinear inverse problems (JPEG artifact removal, nonlinear deblurring, phase
retrieval), all with additive Gaussian noise σ = 0.05, followed by MRI reconstruction with inherently noisy measurements. Top to
bottom: reference, degraded measurement, and PnP-CM reconstruction. Abstract ADMM-based PnP solver that provides a unified frame-
work for solving a wide range of inverse problems, and
incorporates noise perturbations and momentum-based
updates to improve performance in the low-NFE regime.
We evaluate our approach on a diverse set of linear and
nonlinear inverse problems.
We also train and apply
CMs to MRI data for the first time. Our results show that
PnP-CM achieves high-quality reconstructions in as few
as 4 NFEs, and produces meaningful results in 2 steps,
highlighting its effectiveness in real-world inverse problems
while outperforming existing CM-based approaches. Diffusion models have found extensive use in solving in-
verse problems, by sampling from an approximate posterior
distribution of data given the measurements.
Recently,
consistency models (CMs) have been proposed to directly
predict the final output from any point on the diffusion
ODE trajectory, enabling high-quality sampling in just a
few neural function evaluations (NFEs). CMs have also
been utilized for inverse problems, but existing CM-based
solvers either require additional task-specific training
or utilize data fidelity operations with slow convergence,
limiting their applicability to large-scale problems and
making them difficult to extend to nonlinear settings. In
this work, we reinterpret CMs as proximal operators of
a prior, enabling their integration into plug-and-play
(PnP) frameworks. Specifically, we propose PnP-CM, an 1. Introduction Diffusion models (DMs) have established themselves as
state-of-the-art generative models, capable of synthesizing 1",A diagram of the images of people and.,A detailed layout showing a group of people who are smiling and having different faces on them in different colors.,"A comprehensive technical explanation of the differences between multiple personalitys and what they are supposed to be, including the same face of a woman, a man, a woman and a man and a woman with glasses, a couch,.","Figure 1. Representative results of our method (PnP-CM) across a diverse set of inverse problems. Left to right: linear inverse problems
(Gaussian deblurring, super-resolution, inpainting) and nonlinear inverse problems (JPEG artifact removal, nonlinear deblurring, phase
retrieval), all with additive Gaussian noise σ = 0.05, followed by MRI reconstruction with inherently noisy measurements. Top to
bottom: reference, degraded measurement, and PnP-CM reconstruction. Abstract ADMM-based PnP solver that provides a unified frame-
work for solving a wide range of inverse problems, and
incorporates noise perturbations and momentum-based
updates to improve performance in the low-NFE regime.
We evaluate our approach on a diverse set of linear and
nonlinear inverse problems.
We also train and apply
CMs to MRI data for the first time. Our results show that
PnP-CM achieves high-quality reconstructions in as few
as 4 NFEs, and produces meaningful results in 2 steps,
highlighting its effectiveness in real-world inverse problems
while outperforming existing CM-based approaches. Diffusion models have found extensive use in solving in-
verse problems, by sampling from an approximate posterior
distribution of data given the measurements.
Recently,
consistency models (CMs) have been proposed to directly
predict the final output from any point on the diffusion
ODE trajectory, enabling high-quality sampling in just a
few neural function evaluations (NFEs). CMs have also
been utilized for inverse problems, but existing CM-based
solvers either require additional task-specific training
or utilize data fidelity operations with slow convergence,
limiting their applicability to large-scale problems and
making them difficult to extend to nonlinear settings. In
this work, we reinterpret CMs as proximal operators of
a prior, enabling their integration into plug-and-play
(PnP) frameworks. Specifically, we propose PnP-CM, an 1. Introduction Diffusion models (DMs) have established themselves as
state-of-the-art generative models, capable of synthesizing 1",0.75,0.2864,0.5182,4878,2294,2.126,d6e733750237ec476df728c3edf5c05e,images/2025/arxiv_0000050.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000051,Figure 51,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 1. Representative results of our method (PnP-CM) across a diverse set of inverse problems. Left to right: linear inverse problems
(Gaussian deblurring, super-resolution, inpainting) and nonlinear inverse problems (JPEG artifact removal, nonlinear deblurring, phase
retrieval), all with additive Gaussian noise σ = 0.05, followed by MRI reconstruction with inherently noisy measurements. Top to
bottom: reference, degraded measurement, and PnP-CM reconstruction.",A diagram of the four pictures on the.,A detailed layout showing the different photos of people in a photo album and how they are using them.,"A comprehensive technical explanation of the role of the first woman in the u s armed forces in the war of the worlds, including the first female soldier in the united states and second woman in world wars with a world war,.","Figure 1. Representative results of our method (PnP-CM) across a diverse set of inverse problems. Left to right: linear inverse problems
(Gaussian deblurring, super-resolution, inpainting) and nonlinear inverse problems (JPEG artifact removal, nonlinear deblurring, phase
retrieval), all with additive Gaussian noise σ = 0.05, followed by MRI reconstruction with inherently noisy measurements. Top to
bottom: reference, degraded measurement, and PnP-CM reconstruction.",0.7143,0.227,0.4707,1224,897,1.365,757ac855d3bed0d814a71a7cc4dfc65b,images/2025/arxiv_0000051.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000052,Figure 52,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 2. Representative results for Gaussian deblurring, inpainting (70%), and super-resolution (×4). PnP-CM produces sharp and
coherent reconstructions, preserving fine details while avoiding the over-smoothing observed in DPS. Compared to CM-based methods, it
more reliably recovers structured content, and achieves visual quality comparable to recent DM-based PnP methods (e.g., DiffPIR, DPnP,
PnP-DM) while requiring substantially fewer iterations. details are provided in SuppMat B.3. that CM-based approaches often fail to recover. The re-
constructions also preserve high frequency details without
introducing the grid-like artifacts seen in some of the com-
peting methods. Compared to recent DM-based PnP meth-
ods, which can achieve high-quality reconstructions but typ-
ically require substantially more iterations, PnP-CM attains
comparable visual quality while operating at a fraction of
the computational cost.
For nonlinear inverse problems
(Fig. 3), which are intrinsically difficult to solve, PnP-CM
yields reconstructions that are either improved or compa-
rable to those of DPS, ΠGDM, and DPnP, while requiring
significantly fewer NFEs. These visual observations are re-
flected quantitatively in Tab. 1, where PnP-CM reaches state
of the art quality with only a few evaluations. Additional
qualitative comparisons are provided in SuppMat D.4. 4.3. Quantitative and Qualitative Results Given the differing requirements of natural and medical im-
age tasks, we employ separate evaluation metrics for the
two settings. For natural image tasks, reconstruction qual-
ity is assessed using peak signal-to-noise-ratio (PSNR) and
learned perceptual image patch similarity (LPIPS), while
for medical imaging tasks, we use PSNR and structural sim-
ilarity index (SSIM) as evaluation metrics. Quantitative re-
sults are reported as averages over the validation set, while
qualitative examples are provided to illustrate the visual fi-
delity of the reconstructions. For all tasks, we use N = 4
iterations, except for phase retrieval where we use N = 8
due to its increased difficulty. In SuppMat D.3, we further
report results for N ∈{2, 4, 8} and σy = 0.025 to illustrate
the performance and robustness of the method. MRI reconstruction. Representative MRI reconstructions
for R = 8 are depicted in Fig. 4, showing that our method
substantially reduces blurring artifacts observed in DPS, as
well as structured artifacts present in both DPS and DDS.
CM4IR exhibits residual aliasing and noise amplification in
this scenario, as the back-projection term in its data fidelity,
A†(Ax −y) amounts to x−A†y in multi-coil MRI, where Natural image tasks. Representative results for linear in-
verse problems in Fig. 2 demonstrate that PnP-CM consis-
tently yields sharp and high-quality reconstructions. Com-
pared to DPS, our method avoids oversmoothing while be-
ing significantly faster, and it generates coherent structures 7",A diagram of multiple pictures of a.,"A detailed layout showing multiple images of a woman with various facial expressions and hair styles, including a woman ' s face.","A comprehensive technical explanation of a multiple image of a woman with different makeup looks and facial expressions, all in different colors, from brown to blonde to red and black and green and yellow to blue and white and orange and red.","Figure 2. Representative results for Gaussian deblurring, inpainting (70%), and super-resolution (×4). PnP-CM produces sharp and
coherent reconstructions, preserving fine details while avoiding the over-smoothing observed in DPS. Compared to CM-based methods, it
more reliably recovers structured content, and achieves visual quality comparable to recent DM-based PnP methods (e.g., DiffPIR, DPnP,
PnP-DM) while requiring substantially fewer iterations. details are provided in SuppMat B.3. that CM-based approaches often fail to recover. The re-
constructions also preserve high frequency details without
introducing the grid-like artifacts seen in some of the com-
peting methods. Compared to recent DM-based PnP meth-
ods, which can achieve high-quality reconstructions but typ-
ically require substantially more iterations, PnP-CM attains
comparable visual quality while operating at a fraction of
the computational cost.
For nonlinear inverse problems
(Fig. 3), which are intrinsically difficult to solve, PnP-CM
yields reconstructions that are either improved or compa-
rable to those of DPS, ΠGDM, and DPnP, while requiring
significantly fewer NFEs. These visual observations are re-
flected quantitatively in Tab. 1, where PnP-CM reaches state
of the art quality with only a few evaluations. Additional
qualitative comparisons are provided in SuppMat D.4. 4.3. Quantitative and Qualitative Results Given the differing requirements of natural and medical im-
age tasks, we employ separate evaluation metrics for the
two settings. For natural image tasks, reconstruction qual-
ity is assessed using peak signal-to-noise-ratio (PSNR) and
learned perceptual image patch similarity (LPIPS), while
for medical imaging tasks, we use PSNR and structural sim-
ilarity index (SSIM) as evaluation metrics. Quantitative re-
sults are reported as averages over the validation set, while
qualitative examples are provided to illustrate the visual fi-
delity of the reconstructions. For all tasks, we use N = 4
iterations, except for phase retrieval where we use N = 8
due to its increased difficulty. In SuppMat D.3, we further
report results for N ∈{2, 4, 8} and σy = 0.025 to illustrate
the performance and robustness of the method. MRI reconstruction. Representative MRI reconstructions
for R = 8 are depicted in Fig. 4, showing that our method
substantially reduces blurring artifacts observed in DPS, as
well as structured artifacts present in both DPS and DDS.
CM4IR exhibits residual aliasing and noise amplification in
this scenario, as the back-projection term in its data fidelity,
A†(Ax −y) amounts to x−A†y in multi-coil MRI, where Natural image tasks. Representative results for linear in-
verse problems in Fig. 2 demonstrate that PnP-CM consis-
tently yields sharp and high-quality reconstructions. Com-
pared to DPS, our method avoids oversmoothing while be-
ing significantly faster, and it generates coherent structures 7",0.75,0.2825,0.5162,4203,2545,1.651,511e8727dc6f06244c170f163bfd8238,images/2025/arxiv_0000052.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000053,Figure 53,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 2. Representative results for Gaussian deblurring, inpainting (70%), and super-resolution (×4). PnP-CM produces sharp and
coherent reconstructions, preserving fine details while avoiding the over-smoothing observed in DPS. Compared to CM-based methods, it
more reliably recovers structured content, and achieves visual quality comparable to recent DM-based PnP methods (e.g., DiffPIR, DPnP,
PnP-DM) while requiring substantially fewer iterations.",A diagram of multiple images with.,A detailed layout showing multiple images of a woman with different hair styles and make - up for her.,"A comprehensive technical explanation of a photo montage of a woman with long hair and makeup products on her face, showing multiple stages of the process in the same picture, with a woman ' s face, from the same image.","Figure 2. Representative results for Gaussian deblurring, inpainting (70%), and super-resolution (×4). PnP-CM produces sharp and
coherent reconstructions, preserving fine details while avoiding the over-smoothing observed in DPS. Compared to CM-based methods, it
more reliably recovers structured content, and achieves visual quality comparable to recent DM-based PnP methods (e.g., DiffPIR, DPnP,
PnP-DM) while requiring substantially fewer iterations.",0.7143,0.2983,0.5063,1224,762,1.606,24af191e37270b3f71a6316e3122729f,images/2025/arxiv_0000053.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000054,Figure 54,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 3. Comparison of reconstruction quality for nonlinear for-
ward models. PnP-CM yields sharp and coherent reconstructions,
with quality that is competitive with or improved over existing
methods, highlighting robust performance in challenging nonlin-
ear settings while requiring substantially fewer NFEs. m > n, and A† = (A⊤A)−1A⊤. In other words, the solu-
tion moves toward the linear least squares solution, A†y,
which suffers from substantial aliasing artifacts.
Tab. 1
summarizes the performance of different approaches on the
Coronal PD and PD-FS datasets with acceleration rates of
R = 4 and R = 8. Across all settings, our proposed method
with NFE = 4 consistently outperforms DPS (NFE=1000),
DDS (NFE=100) and CM4IR (NFE=4) in terms of PSNR
and SSIM. Additional visual results, including those at the
more modest R = 4, are provided in SuppMat D.5.",A diagram of a woman ' s face with.,"A detailed layout showing the different angles of a woman ' s face with her hair styled, and the same side by side.","A comprehensive technical explanation of the three photos show the same amount of light in each room and the same person with the same hair style on the other side of the picture, in the same room, the same picture, the room.","Figure 3. Comparison of reconstruction quality for nonlinear for-
ward models. PnP-CM yields sharp and coherent reconstructions,
with quality that is competitive with or improved over existing
methods, highlighting robust performance in challenging nonlin-
ear settings while requiring substantially fewer NFEs. m > n, and A† = (A⊤A)−1A⊤. In other words, the solu-
tion moves toward the linear least squares solution, A†y,
which suffers from substantial aliasing artifacts.
Tab. 1
summarizes the performance of different approaches on the
Coronal PD and PD-FS datasets with acceleration rates of
R = 4 and R = 8. Across all settings, our proposed method
with NFE = 4 consistently outperforms DPS (NFE=1000),
DDS (NFE=100) and CM4IR (NFE=4) in terms of PSNR
and SSIM. Additional visual results, including those at the
more modest R = 4, are provided in SuppMat D.5.",0.75,0.2503,0.5001,2273,990,2.296,cc601a54a13cef035471602d6313c591,images/2025/arxiv_0000054.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000055,Figure 55,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 4. Qualitative comparisons of DPS, DDS, CM4IR, and
PnP-CM. Top: Coronal PD with R = 8. Bottom: Coronal PD-FS
with R = 8. PnP-CM effectively reduces artifacts and blurring
that are not removed by other methods (red and yellow arrows).",A diagram of different stages of knee.,A detailed layout showing the different stages of an ostrich joint with a cross section in the middle.,"A comprehensive technical explanation of the impact of ostexyal and ostexial injuries on the knee, including ostex, osteoplasmos and ostrictalitis, ostremals, ostric, ostraplorosis, ostri representing all details, specifications, and configurations.","Figure 4. Qualitative comparisons of DPS, DDS, CM4IR, and
PnP-CM. Top: Coronal PD with R = 8. Bottom: Coronal PD-FS
with R = 8. PnP-CM effectively reduces artifacts and blurring
that are not removed by other methods (red and yellow arrows).",0.7143,0.2098,0.4621,5016,2573,1.949,2f4062da5cdd5748410e4bb882fe6ed2,images/2025/arxiv_0000055.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000056,Figure 56,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 3. Comparison of reconstruction quality for nonlinear for-
ward models. PnP-CM yields sharp and coherent reconstructions,
with quality that is competitive with or improved over existing
methods, highlighting robust performance in challenging nonlin-
ear settings while requiring substantially fewer NFEs.",A diagram of the different images of a.,"A detailed layout showing multiple images of a woman and man with different expressions, including the same person.","A comprehensive technical explanation of the multiple image editing process for a photo booth or video booth, including the same image, and the same picture in one image, the same photo, different format, with each, each, different.","Figure 3. Comparison of reconstruction quality for nonlinear for-
ward models. PnP-CM yields sharp and coherent reconstructions,
with quality that is competitive with or improved over existing
methods, highlighting robust performance in challenging nonlin-
ear settings while requiring substantially fewer NFEs.",0.75,0.3174,0.5337,553,685,0.807,fc6f29fd65782b16f18e674cc262b6b7,images/2025/arxiv_0000056.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000057,Figure 57,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 4. Qualitative comparisons of DPS, DDS, CM4IR, and
PnP-CM. Top: Coronal PD with R = 8. Bottom: Coronal PD-FS
with R = 8. PnP-CM effectively reduces artifacts and blurring
that are not removed by other methods (red and yellow arrows).",A diagram of multiple images of the.,A detailed layout showing multiple images of the various structures of a human body with no visible organs or tissues.,"A comprehensive technical explanation of the effects of the nephoidic in the right knee, and the effects to the left knee are shown in red and green arrows on the lower left side of the image with the image below.","Figure 4. Qualitative comparisons of DPS, DDS, CM4IR, and
PnP-CM. Top: Coronal PD with R = 8. Bottom: Coronal PD-FS
with R = 8. PnP-CM effectively reduces artifacts and blurring
that are not removed by other methods (red and yellow arrows).",0.75,0.2249,0.4874,553,530,1.043,a2d0f85941558bd8974d6cc9d5459fda,images/2025/arxiv_0000057.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000058,Figure 58,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 5. Absolute change in PSNR and LPIPS when perturbing
the reported PnP-CM hyperparameters. Even with up to 25% vari-
ation, the deviations remain small, indicating strong robustness.",A diagram of the distribution of water.,A detailed layout showing the distribution of the perturation from the data of the system and the distribution.,"A comprehensive technical explanation of the flow rate of the aspirators and aspirants in the following stages of the process, asp106, aspiriumation viab / blprs is perturtion representing all details, specifications, and configurations of the.","Figure 5. Absolute change in PSNR and LPIPS when perturbing
the reported PnP-CM hyperparameters. Even with up to 25% vari-
ation, the deviations remain small, indicating strong robustness.",0.8064,0.2766,0.5415,513,324,1.583,37e63634321cd9ce3d6de9f8f1293efc,images/2025/arxiv_0000058.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000059,Figure 59,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 8. Illustrative 4× super-resolution results on CelebA-HQ with σy = 0.05. PnP-CM restores high-frequency details with improved
clarity and denoising, yielding visually richer reconstructions.",A diagram of many different faces of.,"A detailed layout showing the faces of many people on the red carpet at an event, with multiple pictures of them.","A comprehensive technical explanation of this photograph shows the many faces of people in the photo booth on the red carpet of a red carpet event, including president obama and then obama and the other presidents, barack and clintons, obama.","Figure 8. Illustrative 4× super-resolution results on CelebA-HQ with σy = 0.05. PnP-CM restores high-frequency details with improved
clarity and denoising, yielding visually richer reconstructions.",0.75,0.3458,0.5479,4026,4125,0.976,b3d0181a18c74ad32ab7e9c9aa15e5ee,images/2025/arxiv_0000059.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000060,Figure 60,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 8. Illustrative 4× super-resolution results on CelebA-HQ with σy = 0.05. PnP-CM restores high-frequency details with improved
clarity and denoising, yielding visually richer reconstructions.",A diagram of the many faces of people.,"A detailed layout showing a series of different faces of men and women, including one with blonde hair.","A comprehensive technical explanation of many faces in a collage of people and their faces, including barack obama, michelle obama, donald trump, mello ross, and others, and hillary clinton, and other presidents, all.","Figure 8. Illustrative 4× super-resolution results on CelebA-HQ with σy = 0.05. PnP-CM restores high-frequency details with improved
clarity and denoising, yielding visually richer reconstructions.",0.7179,0.3025,0.5102,1224,1276,0.959,4d02fe8d023deaedcded932cc5835991,images/2025/arxiv_0000060.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000061,Figure 61,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 9. Representative Gaussian deblurring results on LSUN Bedroom with σy = 0.05. Comparisons with all baseline methods show
that PnP-CM restores textures more faithfully and avoids oversmoothing.",A diagram of a church with rows of.,"A detailed layout showing multiple photos of a hotel lobby with floral arrangements on the tables and chairs, and a bench in front.","A comprehensive technical explanation of a video editing process including multiple images and texting options, including multiple frames of the same image, and multiple images of each one of a woman in a man sitting on a chair at a table.","Figure 9. Representative Gaussian deblurring results on LSUN Bedroom with σy = 0.05. Comparisons with all baseline methods show
that PnP-CM restores textures more faithfully and avoids oversmoothing.",0.81,0.2298,0.5199,4026,4125,0.976,03892ee3d7d32c85d612801717b16679,images/2025/arxiv_0000061.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000062,Figure 62,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 9. Representative Gaussian deblurring results on LSUN Bedroom with σy = 0.05. Comparisons with all baseline methods show
that PnP-CM restores textures more faithfully and avoids oversmoothing.",A diagram of a machine in a factory.,"A detailed layout showing rows of rows of empty seats in a church with a clock on the wall that shows key features, attributes, and data points in.","A comprehensive technical explanation of a machine to make a quilter ' s fabric on the machine is shown in multiple pictures with a variety of fabrics and colors of machines in front rows of them, including white and pink and red.","Figure 9. Representative Gaussian deblurring results on LSUN Bedroom with σy = 0.05. Comparisons with all baseline methods show
that PnP-CM restores textures more faithfully and avoids oversmoothing.",0.81,0.2677,0.5389,1224,1276,0.959,2a54917a31140f0125b69e3df225ee9f,images/2025/arxiv_0000062.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000063,Figure 63,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 11. Demonstration of super-resolution (×4) results on LSUN Bedroom with σy = 0.05. Reconstructions are compared against all
baseline methods, with PnP-CM producing sharper details and closer resemblance to the ground truth.",A diagram of multiple images of a.,A detailed layout showing the multiple images of a coffee machine and cups on display at a storefront.,"A comprehensive technical explanation of the process of printing a product in multiple stages, including a print and color scheme and a selection of colors for each product, from different to which one is available or more than one color.","Figure 11. Demonstration of super-resolution (×4) results on LSUN Bedroom with σy = 0.05. Reconstructions are compared against all
baseline methods, with PnP-CM producing sharper details and closer resemblance to the ground truth.",0.7143,0.2775,0.4959,4026,4125,0.976,f51d4254e11160f52dbfe0ffd6ccb9e8,images/2025/arxiv_0000063.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000064,Figure 64,scientific_figure,PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems,arXiv Research Authors,2509.22736v2,eess,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems. arXiv:2509.22736v2,"Figure 11. Demonstration of super-resolution (×4) results on LSUN Bedroom with σy = 0.05. Reconstructions are compared against all
baseline methods, with PnP-CM producing sharper details and closer resemblance to the ground truth.",A diagram of the different machines.,A detailed layout showing the various machines used in the manufacturing process and how they work together to produce products.,"A comprehensive technical explanation of the different types of dyes available in the market for sale on the shelves of a store or retail store, including the color and the machine and the bottles of the mixers and the labels.","Figure 11. Demonstration of super-resolution (×4) results on LSUN Bedroom with σy = 0.05. Reconstructions are compared against all
baseline methods, with PnP-CM producing sharper details and closer resemblance to the ground truth.",0.75,0.218,0.484,1224,1276,0.959,d86a35cc9a6e4c407a685197f2f7e6d0,images/2025/arxiv_0000064.png,https://arxiv.org/pdf/2509.22736v2.pdf
arxiv_0000065,Figure 65,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,"Figure 1: Diagram of the divergence-free NN (dfNN) model architecture and directional guidance
learning strategy (left). The quiver plot (right) shows an on-grid ice flux reconstruction by the dfNN
with directional guidance (best model) for a subset of the experimental region over Byrd Glacier. limit data collection, making AIS modelling challenging, yet scientifically paramount. Studies have
found that unconstrained interpolations of sparse and noisy ice thickness measurements from airborne
radar surveys produce unrealistic behaviour in numerical ice sheet models due to flux divergences
[11], [21] (flux: transport of ice volume per unit width and time). In response to this issue, numerical
inversions [11], [17] and, more recently, ML approaches [12], [13], [22] have been developed to
reduce flux artefacts by constraining ice flux interpolations with mass conservation. Both proposed ML approaches [12], [13] employ physics-informed neural networks (PINNs) [23], a
widely used PIML framework. PINNs incorporate an additional loss term that quantifies violations of
physical principles, to learn not only to fit the training data, but to simultaneously minimise physical
inconsistencies, such as flux artefacts. Although PINNs have the flexibility to incorporate multiple
physical principles at once, as demonstrated in [12] and [13], and are hence broadly adopted, the
penalty-based ‘soft’ constraining does not guarantee physical consistency, which has been found to
hinder generalisation performance in the face of data sparsity and noise [24], [25]. Furthermore,
unstable trade-offs between data fit and physics can also affect convergence [26]. In this work,
we show how lesser known hard-constrained divergence-free NNs (dfNNs, see Fig. 1), rooted in
[24], [26], [27], can be used to model exactly mass-conserving ice flux vector fields. To assess
competing PIML paradigms for modelling divergence-free vector fields, we compare hard-, soft-,
and unconstrained NN models — dfNNs, PINNs, and regular NNs — on ice flux reconstruction
for Byrd Glacier, Antarctica, shown in Fig. 1, evaluating both predictive accuracy and physics
adherence. Lastly, informed by the application context [28], we test two extensions across all three
models: (i) directional guidance, a learning strategy that leverages continent-wide ice velocity
observations from satellites via a directional loss term Ldir to align predicted flow with this directional
information beyond sparse flux observation locations, and (ii) incorporating auxiliary predictors (e.g.,
surface elevation). Reproducible experiments and implementations in PyTorch [29] are available at
https://github.com/kimbente/mass_conservation_on_rails. 2
Exact mass conservation with dfNNs Introducing dfNNs. In this work, we address the problem of learning 2D vector fields with NNs
under the constraint of local mass conservation. Let v : R2 →R2 denote a 2D vector field, expressed
as v(x, y) = (u(x, y), v(x, y)), where u and v are the vector components in x- and y-direction,
respectively. The divergence of a vector field is defined as the sum of the partial derivatives of its
components, ∇· v = ∂u/∂x + ∂v/∂y, which quantifies the local rate of expansion or compression.
In balanced flows, this quantity must equal zero. The first NN model to integrate the divergence-free
constraint into the model architecture was introduced in early work of Kuroe et al. [27]. The proposed
approach, model inclusive learning, leverages the property that the symplectic gradient of a scalar
stream function is by design divergence-free (see [6] for vector calculus background). By training the
NN to predict the stream function rather than the vector components, akin to the change of variables
trick, and attaining the vector components by taking the symplectic gradient in a deterministic
differential step, the network outputs continuous, exactly divergence-free vector fields, i.e. with 2",A diagram of the different phases of.,A detailed layout showing the three phases of an artificial cell membrane and the different types of it that are used to produce.,"A comprehensive technical explanation of the quantum - based magnetic field experiment of an n - - inferetry - v - e - v, and the diff - v v - vf - vg - vd representing all details, specifications, and configurations of the system components in.","Figure 1: Diagram of the divergence-free NN (dfNN) model architecture and directional guidance
learning strategy (left). The quiver plot (right) shows an on-grid ice flux reconstruction by the dfNN
with directional guidance (best model) for a subset of the experimental region over Byrd Glacier. limit data collection, making AIS modelling challenging, yet scientifically paramount. Studies have
found that unconstrained interpolations of sparse and noisy ice thickness measurements from airborne
radar surveys produce unrealistic behaviour in numerical ice sheet models due to flux divergences
[11], [21] (flux: transport of ice volume per unit width and time). In response to this issue, numerical
inversions [11], [17] and, more recently, ML approaches [12], [13], [22] have been developed to
reduce flux artefacts by constraining ice flux interpolations with mass conservation. Both proposed ML approaches [12], [13] employ physics-informed neural networks (PINNs) [23], a
widely used PIML framework. PINNs incorporate an additional loss term that quantifies violations of
physical principles, to learn not only to fit the training data, but to simultaneously minimise physical
inconsistencies, such as flux artefacts. Although PINNs have the flexibility to incorporate multiple
physical principles at once, as demonstrated in [12] and [13], and are hence broadly adopted, the
penalty-based ‘soft’ constraining does not guarantee physical consistency, which has been found to
hinder generalisation performance in the face of data sparsity and noise [24], [25]. Furthermore,
unstable trade-offs between data fit and physics can also affect convergence [26]. In this work,
we show how lesser known hard-constrained divergence-free NNs (dfNNs, see Fig. 1), rooted in
[24], [26], [27], can be used to model exactly mass-conserving ice flux vector fields. To assess
competing PIML paradigms for modelling divergence-free vector fields, we compare hard-, soft-,
and unconstrained NN models — dfNNs, PINNs, and regular NNs — on ice flux reconstruction
for Byrd Glacier, Antarctica, shown in Fig. 1, evaluating both predictive accuracy and physics
adherence. Lastly, informed by the application context [28], we test two extensions across all three
models: (i) directional guidance, a learning strategy that leverages continent-wide ice velocity
observations from satellites via a directional loss term Ldir to align predicted flow with this directional
information beyond sparse flux observation locations, and (ii) incorporating auxiliary predictors (e.g.,
surface elevation). Reproducible experiments and implementations in PyTorch [29] are available at
https://github.com/kimbente/mass_conservation_on_rails. 2
Exact mass conservation with dfNNs Introducing dfNNs. In this work, we address the problem of learning 2D vector fields with NNs
under the constraint of local mass conservation. Let v : R2 →R2 denote a 2D vector field, expressed
as v(x, y) = (u(x, y), v(x, y)), where u and v are the vector components in x- and y-direction,
respectively. The divergence of a vector field is defined as the sum of the partial derivatives of its
components, ∇· v = ∂u/∂x + ∂v/∂y, which quantifies the local rate of expansion or compression.
In balanced flows, this quantity must equal zero. The first NN model to integrate the divergence-free
constraint into the model architecture was introduced in early work of Kuroe et al. [27]. The proposed
approach, model inclusive learning, leverages the property that the symplectic gradient of a scalar
stream function is by design divergence-free (see [6] for vector calculus background). By training the
NN to predict the stream function rather than the vector components, akin to the change of variables
trick, and attaining the vector components by taking the symplectic gradient in a deterministic
differential step, the network outputs continuous, exactly divergence-free vector fields, i.e. with 2",0.75,0.2459,0.498,3190,1072,2.976,53b91513a3a9cc7f77e348e18a8253f2,images/2025/arxiv_0000065.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000066,Figure 66,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,"Figure 1: Diagram of the divergence-free NN (dfNN) model architecture and directional guidance
learning strategy (left). The quiver plot (right) shows an on-grid ice flux reconstruction by the dfNN
with directional guidance (best model) for a subset of the experimental region over Byrd Glacier.",A diagram of the structure of a cell.,A detailed layout showing the process of an invertitotor to produce a cell phone battery using a cellphone.,"A comprehensive technical explanation of the structure and function of an ion - activated cell phone, including the location of the battery and the potential energy source for it to operate it inverte to charge it to run it,.","Figure 1: Diagram of the divergence-free NN (dfNN) model architecture and directional guidance
learning strategy (left). The quiver plot (right) shows an on-grid ice flux reconstruction by the dfNN
with directional guidance (best model) for a subset of the experimental region over Byrd Glacier.",0.7321,0.2268,0.4794,1224,414,2.957,e13a64bcb8cb4de05564a0a149be1a3a,images/2025/arxiv_0000066.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000067,Figure 67,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,"Figure 2: Test RMSE (↓) comparison across all model variants, averaged over five runs. Boxed values
indicate mean RMSE, with error bars showing ± std. MAD (top) denotes the Mean Absolute Diver-
gence. dfNNs (proposed, in bold) outperform PINNs & NNs, while directional guidance (proposed,
in bold) improves all models and yields the best-performing variant, dfNN + dir (underlined).",A diagram of the number of different.,"A detailed layout showing the number of different hb - 2s and hb2s in each cell that shows key features, attributes, and data points in detail for.","A comprehensive technical explanation of the number and type of cases with the same amount of cases per year, by type of case, as described in this figure 1, from the study of the data table 2, 2011 - 3.","Figure 2: Test RMSE (↓) comparison across all model variants, averaged over five runs. Boxed values
indicate mean RMSE, with error bars showing ± std. MAD (top) denotes the Mean Absolute Diver-
gence. dfNNs (proposed, in bold) outperform PINNs & NNs, while directional guidance (proposed,
in bold) improves all models and yields the best-performing variant, dfNN + dir (underlined).",0.81,0.2739,0.542,3780,933,4.051,ca009c919487262189af784df58ab12e,images/2025/arxiv_0000067.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000068,Figure 68,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,Figure 3: Gridded predictions of models + dir (best variant per model) for a small test region (white).,A diagram of the different phases of.,"A detailed layout showing the different phases of an air - plane flight from the earth to the moon that shows key features, attributes, and data.","A comprehensive technical explanation of the different types of wind patterns in the atmosphere of earth ' s surface, including wind speed, wind flow, and air pressure, and temperature, and water pressure, in the earth ' nn.",Figure 3: Gridded predictions of models + dir (best variant per model) for a small test region (white).,0.81,0.2277,0.5189,3174,1087,2.92,18205e98d78bb9230048f37c3353cad7,images/2025/arxiv_0000068.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000069,Figure 69,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,"Figure 2: Test RMSE (↓) comparison across all model variants, averaged over five runs. Boxed values
indicate mean RMSE, with error bars showing ± std. MAD (top) denotes the Mean Absolute Diver-
gence. dfNNs (proposed, in bold) outperform PINNs & NNs, while directional guidance (proposed,
in bold) improves all models and yields the best-performing variant, dfNN + dir (underlined).",A diagram of a bar chart with the.,"A detailed layout showing the distribution of the number of different cell phones in the world, including the number.","A comprehensive technical explanation of the number of patients who have been treated for the same type of treatment as well as their age and weight, as per table numbers of patients with a total and age, in the number,.","Figure 2: Test RMSE (↓) comparison across all model variants, averaged over five runs. Boxed values
indicate mean RMSE, with error bars showing ± std. MAD (top) denotes the Mean Absolute Diver-
gence. dfNNs (proposed, in bold) outperform PINNs & NNs, while directional guidance (proposed,
in bold) improves all models and yields the best-performing variant, dfNN + dir (underlined).",0.81,0.2426,0.5263,1224,343,3.569,94e3149148806208c40530879eb5073a,images/2025/arxiv_0000069.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000070,Figure 70,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,Figure 3: Gridded predictions of models + dir (best variant per model) for a small test region (white).,A diagram of the two different types.,"A detailed layout showing an image of the same number of different elements in the model, each with its own.","A comprehensive technical explanation of the new and old model for pnm - sn9, which is being developed by the company in the united states of new york and the us, with the same company in 1994 representing all details, specifications, and.",Figure 3: Gridded predictions of models + dir (best variant per model) for a small test region (white).,0.7957,0.233,0.5143,1224,605,2.023,0ced72ee31573f44cff0e91c4554d79d,images/2025/arxiv_0000070.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000071,Figure 71,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,"Figure 4: Train-Test chequerboard over Byrd Glacier, Antarctica. Training regions are shown in
grey and test regions are shown in white. Points indicate locations with ice flux observations,
collected mainly through airborne geophysical surveys, highlighting the anisotropic nature of the
‘flight line’ data. The Antarctic Polar stereographic coordinate system is used (ESPG:3031, https:
//epsg.io/3031). The selected region (purple frame) corresponds to the region shown in Fig. 3.",A diagram of the location of an area.,"A detailed layout showing the area of a city with many dots and squares in blue, and a small square.","A comprehensive technical explanation of the data visual for the solar system in the city of washington, dc, and washington d c, and the us states of america, and canada, and europe, with a graphed region labeled.","Figure 4: Train-Test chequerboard over Byrd Glacier, Antarctica. Training regions are shown in
grey and test regions are shown in white. Points indicate locations with ice flux observations,
collected mainly through airborne geophysical surveys, highlighting the anisotropic nature of the
‘flight line’ data. The Antarctic Polar stereographic coordinate system is used (ESPG:3031, https:
//epsg.io/3031). The selected region (purple frame) corresponds to the region shown in Fig. 3.",0.7071,0.2426,0.4748,2072,2003,1.034,5eb534c9530048519f18560744bb5992,images/2025/arxiv_0000071.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000072,Figure 72,scientific_figure,Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields,arXiv Research Authors,2510.06286v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Mass Conservation on Rails -- Rethinking Physics-Informed Learning of Ice Flow Vector Fields. arXiv:2510.06286v1,"Figure 4: Train-Test chequerboard over Byrd Glacier, Antarctica. Training regions are shown in
grey and test regions are shown in white. Points indicate locations with ice flux observations,
collected mainly through airborne geophysical surveys, highlighting the anisotropic nature of the
‘flight line’ data. The Antarctic Polar stereographic coordinate system is used (ESPG:3031, https:
//epsg.io/3031). The selected region (purple frame) corresponds to the region shown in Fig. 3.",A diagram of the intersection of two.,A detailed layout showing the location of a network in a city area and a blue line of information about the area.,"A comprehensive technical explanation of the application of the algorithm for using the data in the following image, the data is displayed on the map and the data on the screen below it is the data barcoded as described by the data.","Figure 4: Train-Test chequerboard over Byrd Glacier, Antarctica. Training regions are shown in
grey and test regions are shown in white. Points indicate locations with ice flux observations,
collected mainly through airborne geophysical surveys, highlighting the anisotropic nature of the
‘flight line’ data. The Antarctic Polar stereographic coordinate system is used (ESPG:3031, https:
//epsg.io/3031). The selected region (purple frame) corresponds to the region shown in Fig. 3.",0.81,0.2063,0.5081,1224,1303,0.939,81c02bef2e520d3cf9b2fcafd9ad7867,images/2025/arxiv_0000072.png,https://arxiv.org/pdf/2510.06286v1.pdf
arxiv_0000073,Figure 73,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 1. (a) Darkfield (DFI) image (b) Attenuation (ATTN) image. In the attenuation image, several regions of 
the lungs are partially obscured by the cardiac volume and other organs, whereas in the DFI image these areas 
remain visible as regions of reduce, but still detectable small-angle scattering intensity. In particular the ribs or 
cardiac regions emit little or no scatter compared to the porous lungs. (c) and (d) show the contouring performed 
on the DFI image, with the ATTN image on the side to assist in regions where the lungs are clearly visible in it. Realistic tumor insertion Multiple non-overlapping synthetic tumors were inserted into both the 
left and right lung regions of mouse images (segmented by the above step in the dark-field 
images), based on their respective lung masks. Tumors were placed in identical locations in 
attenuation and dark-field images to maintain channel correspondence.  
The central intensity of each tumor is adjusted for attenuation and darkfield, following expected 
behavior of attenuation and dark-field contrast. Lung tumors have higher attenuation than",A diagram of the diaploid and dfi of.,A detailed layout showing the diaploid of the chest before surgery with the dfi arrows pointing to the abdomen.,"A comprehensive technical explanation of the dfi on the chest and abdomen of a patient with no surgery, including the dft on the abdomen of the chest, and the dfl on the left side of the abdomen, is shown.","Figure 1. (a) Darkfield (DFI) image (b) Attenuation (ATTN) image. In the attenuation image, several regions of 
the lungs are partially obscured by the cardiac volume and other organs, whereas in the DFI image these areas 
remain visible as regions of reduce, but still detectable small-angle scattering intensity. In particular the ribs or 
cardiac regions emit little or no scatter compared to the porous lungs. (c) and (d) show the contouring performed 
on the DFI image, with the ATTN image on the side to assist in regions where the lungs are clearly visible in it. Realistic tumor insertion Multiple non-overlapping synthetic tumors were inserted into both the 
left and right lung regions of mouse images (segmented by the above step in the dark-field 
images), based on their respective lung masks. Tumors were placed in identical locations in 
attenuation and dark-field images to maintain channel correspondence.  
The central intensity of each tumor is adjusted for attenuation and darkfield, following expected 
behavior of attenuation and dark-field contrast. Lung tumors have higher attenuation than",0.7464,0.2979,0.5222,519,555,0.935,f75d162f0215b864f779d234ba0061fd,images/2025/arxiv_0000073.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000074,Figure 74,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,NULL,A diagram of the chest and abdomen.,"A detailed layout showing the position of the rib cage, which is not visible for all of the radiologists to see.","A comprehensive technical explanation of the anatomy of the human chest and abdomens, including the thorble, the thorbla, and the ribs of the abdomens and the abdomen, from the front view of the chest and the upper part of the hip.",NULL,0.75,0.3228,0.5364,524,555,0.944,17100324f67d5cdbcb86a0c4dd3104e1,images/2025/arxiv_0000074.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000075,Figure 75,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,and,A diagram of an x - ray image of a man.,"A detailed layout showing the different areas of the chest and the posterior of the abdomen, including the breast.","A comprehensive technical explanation of the radiology of the chest and abdomen, showing the various areas of the human body with a single breast in the middle section and a three sections on the upper part of the front of the top.",and,0.81,0.2491,0.5295,519,555,0.935,18924c6d23bb1501c0555a73b3c2327e,images/2025/arxiv_0000075.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000076,Figure 76,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,Figure 1.,A diagram of the chest and chest area.,"A detailed layout showing the ribs and chest of an infant with a broken rib - xray image that shows key features, attributes, and data points in.","A comprehensive technical explanation of the location of the chest radiography showing the location and location of a chest radiograph, including the ribs and the chest, the top of a man with a chest, in the chest and an x - ray.",Figure 1.,0.81,0.3476,0.5788,519,555,0.935,ad09c5f15ec004a4773be64e824cb292,images/2025/arxiv_0000076.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000077,Figure 77,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 1. (a) Darkfield (DFI) image (b) Attenuation (ATTN) image. In the attenuation image, several regions of 
the lungs are partially obscured by the cardiac volume and other organs, whereas in the DFI image these areas 
remain visible as regions of reduce, but still detectable small-angle scattering intensity. In particular the ribs or 
cardiac regions emit little or no scatter compared to the porous lungs. (c) and (d) show the contouring performed 
on the DFI image, with the ATTN image on the side to assist in regions where the lungs are clearly visible in it.",A diagram of the chest showing the.,"A detailed layout showing the location of a chest, with a large amount of small areas in the abdomen.","A comprehensive technical explanation of the anatomy of the chest and the lungs, including the thorbone and chest area areas, from the study of the study, c - fyc / p - t, e - l - l representing all details, specifications, and configurations of.","Figure 1. (a) Darkfield (DFI) image (b) Attenuation (ATTN) image. In the attenuation image, several regions of 
the lungs are partially obscured by the cardiac volume and other organs, whereas in the DFI image these areas 
remain visible as regions of reduce, but still detectable small-angle scattering intensity. In particular the ribs or 
cardiac regions emit little or no scatter compared to the porous lungs. (c) and (d) show the contouring performed 
on the DFI image, with the ATTN image on the side to assist in regions where the lungs are clearly visible in it.",0.7107,0.2777,0.4942,976,944,1.034,c9b114c2b4ba27986cc301f97cbe7c04,images/2025/arxiv_0000077.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000078,Figure 78,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 2. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) Corresponding 
contoured dark-field (DFI) image with tumors, where the lesions appear as localized reductions in small-angle 
scattering. (c) Original dark-field image shown for reference.",A diagram of the chest showing an area.,A detailed layout showing the location of the x - rays and the location on the left side of the chest.,"A comprehensive technical explanation of the impact of the presence of the small dots on the chest and the large dots on a large dot on the upper part of the lower part of a large bone, on the right side of the abdomen.","Figure 2. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) Corresponding 
contoured dark-field (DFI) image with tumors, where the lesions appear as localized reductions in small-angle 
scattering. (c) Original dark-field image shown for reference.",0.7143,0.2928,0.5036,980,640,1.531,9887a1ac48f3b937bbd64dff8b9fb0db,images/2025/arxiv_0000078.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000079,Figure 79,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,Original dark-field image shown for reference.,A diagram of the interior of an.,"A detailed layout showing the cat standing on the rock looking down at the camera, in black and white.","A comprehensive technical explanation of the human body ' s skeletal muscles and muscles, including the ribs, legs, and abdomens, and the upper limb bones and the hip bones, from the top view of the posterior view.",Original dark-field image shown for reference.,0.7143,0.2481,0.4812,469,640,0.733,dac6dccba88d7bb4d2586c0c128417e8,images/2025/arxiv_0000079.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000080,Figure 80,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 3. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",A diagram of the chest and ribs shows.,"A detailed layout showing the presence of the metacline on the right side of the chest, and an x - ray image of the left side of a man ' s chest.","A comprehensive technical explanation of the presence of the metaclost on the thorric ribs of a patient in this case, the thorrotrotrotormaploidia is a non - thorromal metator and metatriized representing all details, specifications, and.","Figure 3. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",0.75,0.2918,0.5209,971,700,1.387,943b2517757b7810a5de9125fe512fed,images/2025/arxiv_0000080.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000081,Figure 81,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 3. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",A diagram of a plane flying over a.,A detailed layout showing the dark gray background and the white light on the left side of the image.,"A comprehensive technical explanation of the structure of a moon ' s face and its shadow, including the moon ' ' s tail, which is pointed toward to the left at the sun ' s center of the moon, and the moon.","Figure 3. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",0.7107,0.2651,0.4879,461,700,0.659,e9e991bdb7a4602df1ffc41eac886dd3,images/2025/arxiv_0000081.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000082,Figure 82,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 2. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) Corresponding 
contoured dark-field (DFI) image with tumors, where the lesions appear as localized reductions in small-angle 
scattering. (c) Original dark-field image shown for reference.",A diagram of the different areas of.,"A detailed layout showing the location of the left shoulder and the right shoulder bone, and the patient ' s chest.","A comprehensive technical explanation of the bone marrow and its role in surgery for the hip joint, including the pelvic region and the lumbas of the pelvis bone verteus, for the pectoral bone representing all details, specifications, and.","Figure 2. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) Corresponding 
contoured dark-field (DFI) image with tumors, where the lesions appear as localized reductions in small-angle 
scattering. (c) Original dark-field image shown for reference.",0.75,0.2282,0.4891,987,437,2.259,b1d87ac95aa22b481f169b9f28adb3ed,images/2025/arxiv_0000082.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000083,Figure 83,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 3. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",A diagram of the normal and a normal.,"A detailed layout showing the presence of the papii on the left side of the chest, and the location of the peritoidus on the right side.","A comprehensive technical explanation of x - ray images showing the location and function of the bone, including the end of the right side of the thor and the thorcrasm, and the upper thorblex, and upper.","Figure 3. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly higher 
attenuation than the surrounding lung tissue, making them visible in the upper lung regions but nearly 
indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",0.75,0.2463,0.4981,987,485,2.035,c25df9053237336407ce267775d20a5e,images/2025/arxiv_0000083.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000084,Figure 84,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 4. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly 
higher attenuation than the surrounding lung tissue, making them visible in the upper lung regions but 
nearly indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference. UNET architecture Training/Testing 
2-channel UNET (ATTN and DFI): A two-dimensional U-Net architecture was used for patch-
based tumor segmentation using dual-channel input comprising attenuation (ATTN) and dark-
field (DFI) images. Each 32×32 patch for ATTN and corresponding patch for DFI was fed as a 
3D tensor of dimension [H x W x C], where H=W=32 and C=2, to jointly exploit the contrast 
from both imaging modalities. The encoder consisted of two levels, each containing two 3×3 
convolutional layers with ReLU activations followed by 2×2 max pooling. The number of 
feature channels doubled at each down-sampling stage (16–32–64). The decoder mirrored the 
encoder using 2×2 transposed convolutions for up-sampling and skip connections from 
corresponding encoder layers to preserve spatial detail. A 1×1 convolution generated two output 
logits (background and tumor), which were converted to foreground probabilities via a sigmoid 
or softmax activation. 
Training loss was a hybrid binary cross-entropy + Dice loss, Adam optimizer (learning rate = 
5×10⁻⁴, β₁ = 0.9, β₂ = 0.999), global-norm gradient clipping (1.0), and early stopping based on 
validation loss. Random flips and 90° rotations were used for data augmentation. 1-channel UNET (DFI or Attn): Two single-input variant of the same U-Net architecture were 
trained using only the dark-field (DFI) image or the attenuation (ATTN) channel as input. The",A diagram of the head of a man with a.,"A detailed layout showing the location of the left vental in the upper part of the abdomen, and the yellow line on the lower part of an x - ray.",A comprehensive technical explanation of the chest radiograph of a person with a proxitoid on the right side of the body and a yellow line indicating a large area on the left side of his chest radiogram in the chest.,"Figure 4. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly 
higher attenuation than the surrounding lung tissue, making them visible in the upper lung regions but 
nearly indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference. UNET architecture Training/Testing 
2-channel UNET (ATTN and DFI): A two-dimensional U-Net architecture was used for patch-
based tumor segmentation using dual-channel input comprising attenuation (ATTN) and dark-
field (DFI) images. Each 32×32 patch for ATTN and corresponding patch for DFI was fed as a 
3D tensor of dimension [H x W x C], where H=W=32 and C=2, to jointly exploit the contrast 
from both imaging modalities. The encoder consisted of two levels, each containing two 3×3 
convolutional layers with ReLU activations followed by 2×2 max pooling. The number of 
feature channels doubled at each down-sampling stage (16–32–64). The decoder mirrored the 
encoder using 2×2 transposed convolutions for up-sampling and skip connections from 
corresponding encoder layers to preserve spatial detail. A 1×1 convolution generated two output 
logits (background and tumor), which were converted to foreground probabilities via a sigmoid 
or softmax activation. 
Training loss was a hybrid binary cross-entropy + Dice loss, Adam optimizer (learning rate = 
5×10⁻⁴, β₁ = 0.9, β₂ = 0.999), global-norm gradient clipping (1.0), and early stopping based on 
validation loss. Random flips and 90° rotations were used for data augmentation. 1-channel UNET (DFI or Attn): Two single-input variant of the same U-Net architecture were 
trained using only the dark-field (DFI) image or the attenuation (ATTN) channel as input. The",0.75,0.3253,0.5376,432,715,0.604,3f3de7716feb839de1a2cc3e74ff6a75,images/2025/arxiv_0000084.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000085,Figure 85,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,NULL,A diagram of the presence of a cell in.,"A detailed layout showing the structure of the cell membrane of the microtubr, which is a layer that shows key features, attributes, and data points.","A comprehensive technical explanation of the structure of the human cell membranes and its connections to the cell wall and surrounding the cell, from the surface of the cell line of the plate, which is shown in yellow lines on the image.",NULL,0.87,0.2368,0.5534,435,715,0.608,5bf83da991feb9ad3f252a2df520b0aa,images/2025/arxiv_0000085.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000086,Figure 86,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,NULL,A diagram of a cat ' s head showing.,"A detailed layout showing the light from a microscope at night, and in the dark of the image that shows key features, attributes, and data points in.","A comprehensive technical explanation of the first ever image of a horse in the dark, with a white spot on its forehead and a black background to it ' s left side and right of the head, as well as well,.",NULL,0.81,0.2609,0.5354,437,715,0.611,e70a04a1202009a79fcc4cb0f89b48fe,images/2025/arxiv_0000086.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000087,Figure 87,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 4. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly 
higher attenuation than the surrounding lung tissue, making them visible in the upper lung regions but 
nearly indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",A diagram of the chest and chest area.,A detailed layout showing the location of the left upper lung and the right lower lung of the upper lung.,"A comprehensive technical explanation of the location and function of the spot on the chest, including the location of the small dots in the upper right side of the chest and lower section of the abdomen, and the upper part of the lower.","Figure 4. (a) Contoured attenuation (ATTN) image with inserted tumors. The tumors exhibit slightly 
higher attenuation than the surrounding lung tissue, making them visible in the upper lung regions but 
nearly indistinguishable in the lower sections, where overlapping organs project higher attenuation. (b) 
Corresponding contoured dark-field (DFI) image with tumors, where the lesions appear as localized 
reductions in small-angle scattering. (c) Original dark-field image shown for reference.",0.725,0.2326,0.4788,904,473,1.911,0ffa8cd10a873a2cc4f3961020b94220,images/2025/arxiv_0000087.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000088,Figure 88,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 5. For the single-channel ATTN-only 
UNET model, the training and validation losses are 
shown across epochs. The maximum epoch count 
was set to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",A diagram of a graph showing the.,A detailed layout showing the training vs validation loss and the training variation of each model in the graph.,A comprehensive technical explanation of training vs valuation loss in the u s and canada ' s national train and track association - epochch chart - 1 / 3 / 2 / 3 - 6 / 5 - 4 / 3k / 8 / 4 / 20 / 14 / 12 / 15 / 11.,"Figure 5. For the single-channel ATTN-only 
UNET model, the training and validation losses are 
shown across epochs. The maximum epoch count 
was set to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",0.99,0.3436,0.6668,574,390,1.472,f86fe346632f544b52bcf60195b946c4,images/2025/arxiv_0000088.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000089,Figure 89,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 6. For the single-channel DFI-only UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",A diagram of the training vs depicting.,"A detailed layout showing the training and validation loss of each course in the course, as well as the following.","A comprehensive technical explanation of training vs validation loss in excel and epocht - based data systems - part 3 - data visual, graphics, and visual, text, and design, v3 0, 2013, pk representing all details, specifications, and.","Figure 6. For the single-channel DFI-only UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",0.93,0.2883,0.6092,574,392,1.464,ed6be574faec0756913fc5b5a0518c49,images/2025/arxiv_0000089.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000090,Figure 90,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 5. For the single-channel ATTN-only 
UNET model, the training and validation losses are 
shown across epochs. The maximum epoch count 
was set to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",A diagram of the graph shows the.,"A detailed layout showing the differences between training and validation loss in the same time of the course, each with a different plot.","A comprehensive technical explanation of training vs validation loss in the epoth system and how to use it? - speqch blog - epop com - png - pk - p1 png png, png representing all details, specifications, and configurations of the system components.","Figure 5. For the single-channel ATTN-only 
UNET model, the training and validation losses are 
shown across epochs. The maximum epoch count 
was set to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",0.99,0.3444,0.6672,493,326,1.512,f08b49f93cf29031cf300a55baacd082,images/2025/arxiv_0000090.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000091,Figure 91,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 6. For the single-channel DFI-only UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",A diagram of the training and.,"A detailed layout showing the path of training and validation loss in each region of the course, including the following.",A comprehensive technical explanation of the impact of training validation loss in epocht and epochts in the development of an ecicht - based medical system with a single - erp - based approach and ecich therapy.,"Figure 6. For the single-channel DFI-only UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",0.93,0.2776,0.6038,489,332,1.473,49488261cf56eca37587099b6fed0068,images/2025/arxiv_0000091.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000092,Figure 92,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 7. For the two-channel ATTN+DFI UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting. The training results of the single-channel UNET with ATTN-only and DFI-only are shown in 
Fig. 5-6. Fig. 7 shows that of 2-channel architecture of ATTN+DFI. The Table 1 shows that the 
true-positive sensitivity improved with dark-field only patches to 83.75 from 51% with just 
attenuation patches. The Specificity was slightly better with attenuation 92.9% (ATTN-only) 
versus 90.5% with DFI-ONLY. The ATTN+DFI has intermediate results of Sensitivity of 79.6% 
and improved Specificity to 97.6%. Attn    Pred    GT          DFI    Pred   GT        ATTN+DFI    Pred   GT",A diagram of the training vs depicting.,A detailed layout showing the training vs validation loss versus the training / validation loss graph in excel and excel.,"A comprehensive technical explanation of training vs validation loss and the path to completion of each training session in a single file, a plot is shown, from the data source of the following the graphing process in the above the chart.","Figure 7. For the two-channel ATTN+DFI UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting. The training results of the single-channel UNET with ATTN-only and DFI-only are shown in 
Fig. 5-6. Fig. 7 shows that of 2-channel architecture of ATTN+DFI. The Table 1 shows that the 
true-positive sensitivity improved with dark-field only patches to 83.75 from 51% with just 
attenuation patches. The Specificity was slightly better with attenuation 92.9% (ATTN-only) 
versus 90.5% with DFI-ONLY. The ATTN+DFI has intermediate results of Sensitivity of 79.6% 
and improved Specificity to 97.6%. Attn    Pred    GT          DFI    Pred   GT        ATTN+DFI    Pred   GT",0.99,0.3359,0.6629,574,393,1.461,569322f87d2e536a20888cae35422da8,images/2025/arxiv_0000092.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000093,Figure 93,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 7. For the two-channel ATTN+DFI UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",A diagram of the training and.,A detailed layout showing the training vs validation loss plot and the results of each test on the same line.,"A comprehensive technical explanation of training vs validation loss and time of training in the same class of training period, based on the following steps in each step, the model, the graphed - up to the next slide - click.","Figure 7. For the two-channel ATTN+DFI UNET 
model, the training and validation losses are shown 
across epochs. The maximum epoch count was set 
to 100, but the training stopped early due to 
oscillations in the validation loss, possibly 
indicating the onset of overfitting.",0.9793,0.3538,0.6665,490,329,1.489,5471460153bce2fca174782720d67f9e,images/2025/arxiv_0000093.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000094,Figure 94,scientific_figure,Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models,arXiv Research Authors,2510.27679v1,physics,2025,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2025). Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models. arXiv:2510.27679v1,"Figure 8. The test panels show the patch(es), predicted mask and the ground truth (GT) masks 
for Attn-only, DFI-only and ATTN+DFI UNET results. The top and bottom panels show 
patches where the ATTN-only misses a tumor but DFI-only or ATTN+DFI finds the tumor.",A diagram of the structure of a cell.,"A detailed layout showing the number of possible samples for the first sample of atn - nth that shows key features, attributes, and data points in.","A comprehensive technical explanation of the first - gen gen trial for atfd1, which is the most effective and most advanced to treat the condition of the disease of atfgdg all - gen - gen 1 representing all details, specifications, and.","Figure 8. The test panels show the patch(es), predicted mask and the ground truth (GT) masks 
for Attn-only, DFI-only and ATTN+DFI UNET results. The top and bottom panels show 
patches where the ATTN-only misses a tumor but DFI-only or ATTN+DFI finds the tumor.",0.81,0.3184,0.5642,928,539,1.722,8bf7779b2988afda3d0b2c07e0071dd1,images/2025/arxiv_0000094.png,https://arxiv.org/pdf/2510.27679v1.pdf
arxiv_0000095,Figure 95,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 1. Reconstruction comparison across probe types and acquisition modes. Rows:
idealized probe (Gaussian-smoothed disk, uniform phase) vs semi-synthetic (experi-
mental probe, synthetic object). Columns: single-shot CDI vs overlapped ptychography.",A diagram of a yellow and blue square.,"A detailed layout showing the size of the pixel pattern in which the lines appear to be cut out that shows key features, attributes, and data points.","A comprehensive technical explanation of the heat map for this solar system, including the radiation spectrum and the current temperature of the sun in the atmosphere of the earth ' s atmosphere, as shown in the same direction by the sun ' s.","Fig. 1. Reconstruction comparison across probe types and acquisition modes. Rows:
idealized probe (Gaussian-smoothed disk, uniform phase) vs semi-synthetic (experi-
mental probe, synthetic object). Columns: single-shot CDI vs overlapped ptychography.",0.81,0.2416,0.5258,800,798,1.003,40f5a7ebfcf6befdc002b211a03d1389,images/2026/arxiv_0000095.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000096,Figure 96,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,Idealized — Ptycho,A diagram of the cell in the human.,A detailed layout showing the effects of a cell phone radiation image with a star - like pattern in blue and yellow.,"A comprehensive technical explanation of the human cell structure and its functions in the body, from the study of homo - based cell structures to the cellular structure of the cell wall of the brain and cell membranes, and the cell.",Idealized — Ptycho,0.75,0.3272,0.5386,800,800,1.0,99d8852b692085560dd9cd7a58a1a555,images/2026/arxiv_0000096.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000097,Figure 97,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,Semi-synthetic — CDI,A diagram of the cell membranes in the.,A detailed layout showing the location of the cell phone and the location where it is located in the image.,"A comprehensive technical explanation of the study of the cell wall in a cell phone case by dr j jl, m d, and dr jl m s d, m, mds w c, md, r s representing all details, specifications, and configurations of the system components in full context.",Semi-synthetic — CDI,0.7321,0.2231,0.4776,800,797,1.004,4decc0110aecb68eee8713562a5ca706,images/2026/arxiv_0000097.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000098,Figure 98,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,NULL,A diagram of the cell membranes in an.,A detailed layout showing the structure of the human cell membranes and the areas where each cell is located.,"A comprehensive technical explanation of the presence of a cell in an area of the body of matter, and the potential of a new generation of cells that are present in this body - cell biology, a human, is a cell.",NULL,0.7393,0.2379,0.4886,800,798,1.003,0ebaa3c4c146191d00050768d7853af3,images/2026/arxiv_0000098.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000099,Figure 99,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 1. Reconstruction comparison across probe types and acquisition modes. Rows:
idealized probe (Gaussian-smoothed disk, uniform phase) vs semi-synthetic (experi-
mental probe, synthetic object). Columns: single-shot CDI vs overlapped ptychography.",A diagram of the different types of.,A detailed layout showing the effects of the sun ' s radiation on the surface of earth and in space.,"A comprehensive technical explanation of the heat map of the celluloon - based model of the human body, showing the various areas of the tissue structure of the cells and the tissues and the cell bodies in the cells,.","Fig. 1. Reconstruction comparison across probe types and acquisition modes. Rows:
idealized probe (Gaussian-smoothed disk, uniform phase) vs semi-synthetic (experi-
mental probe, synthetic object). Columns: single-shot CDI vs overlapped ptychography.",0.7071,0.2044,0.4557,695,1006,0.691,65b768b35b56bd21b4f84bc171e6c4a2,images/2026/arxiv_0000099.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000100,Figure 100,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 2. Comparison of reconstruction quality with different numbers of diffraction
patterns.",A diagram of the different images are.,"A detailed layout showing the different patterns of optical distortion in the image, including the sunburst.","A comprehensive technical explanation of optical patterns for the visual effects of motion in motion and time - lapss, including the movement of the eye and the eye, and the focals of the image, and direction of the lens.","Fig. 2. Comparison of reconstruction quality with different numbers of diffraction
patterns.",0.7357,0.2887,0.5122,2048,999,2.05,50016ff4f08ca242ea8283d146288a2d,images/2026/arxiv_0000100.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000101,Figure 101,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 2. Comparison of reconstruction quality with different numbers of diffraction
patterns.",A diagram of the motion of a computer.,A detailed layout showing the multiple patterns of a sun ray with different colors and sizes in each one.,"A comprehensive technical explanation of the optical image and visual of the image in motion using the optical view of the object, the image is a computerized image, and the image, a standard version of the actual image,.","Fig. 2. Comparison of reconstruction quality with different numbers of diffraction
patterns.",0.725,0.2627,0.4939,692,1304,0.531,f71f2a65fcb11a39b24e6788298a1684,images/2026/arxiv_0000101.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000102,Figure 102,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 3. Photon-limited performance for two self-supervised PtychoPINN variants
trained with mean absolute error (MAE) and Poisson negative log likelihood (NLL)
reconstruction penalties.",A diagram of a graph with a curve and.,"A detailed layout showing the function of exposure data for fsc to a function of exosure dose that shows key features, attributes, and data points.","A comprehensive technical explanation of the fpx - as function of exposure doses for all objects in a single phase and a single output of exposures to one or two phases for each phase, each time, each of the same.","Fig. 3. Photon-limited performance for two self-supervised PtychoPINN variants
trained with mean absolute error (MAE) and Poisson negative log likelihood (NLL)
reconstruction penalties.",0.81,0.3644,0.5872,860,558,1.541,c2424eaa9774c291e753e9829b4a84a1,images/2026/arxiv_0000102.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000103,Figure 103,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 3. Photon-limited performance for two self-supervised PtychoPINN variants
trained with mean absolute error (MAE) and Poisson negative log likelihood (NLL)
reconstruction penalties.",A diagram of a computer screen showing.,"A detailed layout showing the different types of photon and their applications in space, including a large telescope.","A comprehensive technical explanation of the model of photon and photon in the cell phone, with a ploter, and a graphed line depicting the process of the same size for each cell phone number of photons and the cell.","Fig. 3. Photon-limited performance for two self-supervised PtychoPINN variants
trained with mean absolute error (MAE) and Poisson negative log likelihood (NLL)
reconstruction penalties.",0.75,0.2277,0.4889,694,1184,0.586,2ff5497a6eba0b35596dfd9a8a4cc9bf,images/2026/arxiv_0000103.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000104,Figure 104,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 4. Structural similarity of PtychoPINN and the supervised baseline as a function
of training set size.",A diagram of a graph showing the.,"A detailed layout showing the size of the scm phase mean performance versus training set s / t that shows key features, attributes, and data points.","A comprehensive technical explanation of the ssm phase mean performance vs training size for the s - series and s - class models, including s - max and sxpvm - x - max - series, s - v representing all details, specifications, and configurations of.","Fig. 4. Structural similarity of PtychoPINN and the supervised baseline as a function
of training set size.",0.93,0.334,0.632,1840,1164,1.581,ba1ec8c900bcc8ebb77480ec0449e4e5,images/2026/arxiv_0000104.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000105,Figure 105,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 4. Structural similarity of PtychoPINN and the supervised baseline as a function
of training set size.",A diagram of the rise in training time.,"A detailed layout showing the performance of different training levels for each individual individual figure, and the results of each.","A comprehensive technical explanation of the performance of a training set for a woman or man in a competitive sports competition, including the training sets and training sets for men, as well as well, as per the leveling, training.","Fig. 4. Structural similarity of PtychoPINN and the supervised baseline as a function
of training set size.",0.99,0.2898,0.6399,693,588,1.179,2b371a7c0c3c7d8665ad260ff6549b3e,images/2026/arxiv_0000105.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000106,Figure 106,scientific_figure,Towards single-shot coherent imaging via overlap-free ptychography,arXiv Research Authors,2602.21361v3,physics,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). Towards single-shot coherent imaging via overlap-free ptychography. arXiv:2602.21361v3,"Fig. 5. Comparison of methods for an in-distribution LCLS control (train LCLS XPP,
test LCLS XPP) and out-of-distribution transfer (train APS, test LCLS XPP). The
reference column shows an ePIE reconstruction of the LCLS data.",A diagram of the letters n and m are.,"A detailed layout showing the different types of letters in each image and the different colors of them in black, yellow, red, green, orange, and.","A comprehensive technical explanation of the initial and initial letters of the alphabet m, m, n, d, s, e, n and fo, f, f in the image of the same color as shown in the following representing all details, specifications, and configurations of the.","Fig. 5. Comparison of methods for an in-distribution LCLS control (train LCLS XPP,
test LCLS XPP) and out-of-distribution transfer (train APS, test LCLS XPP). The
reference column shows an ePIE reconstruction of the LCLS data.",0.75,0.2837,0.5169,652,508,1.283,4546f4ba684941425b8347cb0b781dc6,images/2026/arxiv_0000106.png,https://arxiv.org/pdf/2602.21361v3.pdf
arxiv_0000107,Figure 107,scientific_figure,MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI,arXiv Research Authors,2604.11762v1,cs,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI. arXiv:2604.11762v1,"Figure 1: MosaicMRI overview. (left) Anatomy distribution by volume count, showing a long-tailed composition prevalently by
spine (49%, 1,316 volumes), followed by shoulder (14%, 373) and knee (14%, 362). (right) Representative slices spanning six
anatomy groups and three orientations (axial, sagittal, coronal); overlays report in-plane matrix size, receive-coil count, and number
of slices.",A diagram of the different types of.,A detailed layout showing the different stages of the human body in black and white photos with a pie chart.,"A comprehensive technical explanation of the anatomy of the human body and its functions, including the major bones, muscles, and muscles of the pel and pelviss, from the body to the knee, for the hip, the hip.","Figure 1: MosaicMRI overview. (left) Anatomy distribution by volume count, showing a long-tailed composition prevalently by
spine (49%, 1,316 volumes), followed by shoulder (14%, 373) and knee (14%, 362). (right) Representative slices spanning six
anatomy groups and three orientations (axial, sagittal, coronal); overlays report in-plane matrix size, receive-coil count, and number
of slices.",0.7957,0.2966,0.5461,976,516,1.891,d2333164f4a1c8ea25161aaea951a46a,images/2026/arxiv_0000107.png,https://arxiv.org/pdf/2604.11762v1.pdf
arxiv_0000108,Figure 108,scientific_figure,MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI,arXiv Research Authors,2604.11762v1,cs,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI. arXiv:2604.11762v1,"Figure 2: PSNR versus training-set fraction for E2E-VarNet on MOSAICMRI. Each curve is evaluated on the corresponding
anatomy-specific test set. (a) E2E-VarNet with 4 cascades, (b) 8 cascades, and (c) 12 cascades.",A diagram of the number of people with.,"A detailed layout showing the number of people who have visited the area in each country, and how many.","A comprehensive technical explanation of the number of people who use internet in the past decade of years, by age and ethnicity, by race and gender, 2010 - based on the same years of birth and age, 2012 to -.","Figure 2: PSNR versus training-set fraction for E2E-VarNet on MOSAICMRI. Each curve is evaluated on the corresponding
anatomy-specific test set. (a) E2E-VarNet with 4 cascades, (b) 8 cascades, and (c) 12 cascades.",0.7179,0.169,0.4435,930,303,3.069,5f3bd1f8a0b588340207e2c711a23a8d,images/2026/arxiv_0000108.png,https://arxiv.org/pdf/2604.11762v1.pdf
arxiv_0000109,Figure 109,scientific_figure,MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI,arXiv Research Authors,2604.11762v1,cs,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI. arXiv:2604.11762v1,"Figure 3: Mean PSNR (dB) of E2E-VarNet for cross-anatomy transfer on MOSAICMRI. Rows are test anatomies, and columns are
the training anatomy; the final Baseline column corresponds to the model trained on all data. Anatomies are ordered from higher
to lower volume counts (top/left to bottom/right), yielding three groups: high-data anchors (blue), distal extremities (brown), and
low-data groups (pink). For each test anatomy (row), black outlines mark all anatomy models within 1 dB of the best anatomy result
for that row (excluding the baseline).",A diagram of the average time of.,"A detailed layout showing the number of people who are selected by the number in each country, including the number.","A comprehensive technical explanation of the average time to visit a city in the united states by region of origin, from the bureau of customs and state governmentss of texas and state of texas, texas, 2007 - 2011 to 2013.","Figure 3: Mean PSNR (dB) of E2E-VarNet for cross-anatomy transfer on MOSAICMRI. Rows are test anatomies, and columns are
the training anatomy; the final Baseline column corresponds to the model trained on all data. Anatomies are ordered from higher
to lower volume counts (top/left to bottom/right), yielding three groups: high-data anchors (blue), distal extremities (brown), and
low-data groups (pink). For each test anatomy (row), black outlines mark all anatomy models within 1 dB of the best anatomy result
for that row (excluding the baseline).",0.75,0.2598,0.5049,602,579,1.04,d1baf8fd1fa6bd1387d7dbb3f44deb70,images/2026/arxiv_0000109.png,https://arxiv.org/pdf/2604.11762v1.pdf
arxiv_0000110,Figure 110,scientific_figure,MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI,arXiv Research Authors,2604.11762v1,cs,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI. arXiv:2604.11762v1,"Figure 4: Protocol generalization in MOSAICMRI. Mean PSNR (dB) for E2E-VarNet trained on each protocol (columns) and tested
on each protocol (rows); Baseline is trained on all protocols. Boxes mark single-protocol models within 1 dB of the best per row.",A diagram of the number of people on.,"A detailed layout showing the number of people in each country who are in the race, and how many numbers are on each side of the table.","A comprehensive technical explanation of the average number of players in a game of tennis, from the top to the bottom of the table, with the score on the field, and the numbers of the first half - time at each.","Figure 4: Protocol generalization in MOSAICMRI. Mean PSNR (dB) for E2E-VarNet trained on each protocol (columns) and tested
on each protocol (rows); Baseline is trained on all protocols. Boxes mark single-protocol models within 1 dB of the best per row.",0.81,0.222,0.516,790,374,2.112,0dd580857a03c96747992fb674151887,images/2026/arxiv_0000110.png,https://arxiv.org/pdf/2604.11762v1.pdf
arxiv_0000111,Figure 111,scientific_figure,MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI,arXiv Research Authors,2604.11762v1,cs,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI. arXiv:2604.11762v1,"Figure 5: Dataset diversity by anatomy. Violin/box plots summarize the distribution of receive-coil counts and slices per volume for
each anatomy group. A contrast panel reports the presence of major contrast families (T1, T1-FS, T2, T2-FS, PD, PD-FS, STIR)
per anatomy, reflecting protocol heterogeneity. A stacked bar chart shows the orientation mix (axial/sagittal/coronal) within each
anatomy, highlighting anatomy-dependent acquisition geometry.",A diagram of multiple colored lines.,"A detailed layout showing the basic color scheme for a computer keyboard keyboard, including a keyboard and several different colored.","A comprehensive technical explanation of the different types of dnas in the human body, including the size and number of each type of dna that is present in the system, including, the dna and the set to which is different.","Figure 5: Dataset diversity by anatomy. Violin/box plots summarize the distribution of receive-coil counts and slices per volume for
each anatomy group. A contrast panel reports the presence of major contrast families (T1, T1-FS, T2, T2-FS, PD, PD-FS, STIR)
per anatomy, reflecting protocol heterogeneity. A stacked bar chart shows the orientation mix (axial/sagittal/coronal) within each
anatomy, highlighting anatomy-dependent acquisition geometry.",0.75,0.2917,0.5209,5297,1292,4.1,fdf5ee0f419b9c73c7a168be730bc4fe,images/2026/arxiv_0000111.png,https://arxiv.org/pdf/2604.11762v1.pdf
arxiv_0000112,Figure 112,scientific_figure,MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI,arXiv Research Authors,2604.11762v1,cs,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI. arXiv:2604.11762v1,"Figure 5: Dataset diversity by anatomy. Violin/box plots summarize the distribution of receive-coil counts and slices per volume for
each anatomy group. A contrast panel reports the presence of major contrast families (T1, T1-FS, T2, T2-FS, PD, PD-FS, STIR)
per anatomy, reflecting protocol heterogeneity. A stacked bar chart shows the orientation mix (axial/sagittal/coronal) within each
anatomy, highlighting anatomy-dependent acquisition geometry.",A diagram of the number of people in.,A detailed layout showing the different types of the data that is generated in the system and how it works.,"A comprehensive technical explanation of the data visual for a project with a single - columned chart and data visual to the other side of the diagram, including the data, and the same data visual, with an additional information,.","Figure 5: Dataset diversity by anatomy. Violin/box plots summarize the distribution of receive-coil counts and slices per volume for
each anatomy group. A contrast panel reports the presence of major contrast families (T1, T1-FS, T2, T2-FS, PD, PD-FS, STIR)
per anatomy, reflecting protocol heterogeneity. A stacked bar chart shows the orientation mix (axial/sagittal/coronal) within each
anatomy, highlighting anatomy-dependent acquisition geometry.",0.7321,0.2643,0.4982,976,805,1.212,64c76d42d1e6bee3c021759cc1d5b9c2,images/2026/arxiv_0000112.png,https://arxiv.org/pdf/2604.11762v1.pdf
arxiv_0000113,Figure 113,scientific_figure,MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI,arXiv Research Authors,2604.11762v1,cs,2026,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2026). MosaicMRI: A Diverse Dataset and Benchmark for Raw Musculoskeletal MRI. arXiv:2604.11762v1,"Figure 6: Qualitative accelerated reconstruction examples across anatomies. For each panel, columns (left to right) show the
masked k-space after applying the undersampling pattern, the zero-filled RSS reconstruction, the reconstruction produced by VarNet
trained on full MOSAICMRI, and the fully sampled target.",A diagram of the different types of.,"A detailed layout showing a variety of images of the human torso and other parts of the body, including the ribs.","A comprehensive technical explanation of the structure of the human stomach and its major muscles and their functions, including the insertions of the ribs and the outer structures of the lower limbiscanum and the pel bones,.","Figure 6: Qualitative accelerated reconstruction examples across anatomies. For each panel, columns (left to right) show the
masked k-space after applying the undersampling pattern, the zero-filled RSS reconstruction, the reconstruction produced by VarNet
trained on full MOSAICMRI, and the fully sampled target.",0.75,0.2402,0.4951,977,391,2.499,15bbf0d2e2e2b5de74dac508435fb985,images/2026/arxiv_0000113.png,https://arxiv.org/pdf/2604.11762v1.pdf
arxiv_0000114,Figure 114,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 1. High-level visualization of our approach to video de-
scription generation.
We incorporate models of both the local
temporal dynamic (i.e. within blocks of a few frames) of videos,
as well as their global temporal structure. The local structure is
modeled using the temporal feature maps of a 3-D CNN, while
a temporal attention mechanism is used to combine information
across the entire video. For each generated word, the model can
focus on different temporal regions in the video. For simplicity, we
highlight only the region having the maximum attention above.",A diagram of three different shots.,"A detailed layout showing the shooting gun and the shooting position of the man in the mirror, as well as four frames with different images.","A comprehensive technical explanation of shooting a gun for beginners, part 2 of 3 - - - the basics and advanced techniques for beginers - photo 1 - click to read here or click to enlargeze the picture.","Figure 1. High-level visualization of our approach to video de-
scription generation.
We incorporate models of both the local
temporal dynamic (i.e. within blocks of a few frames) of videos,
as well as their global temporal structure. The local structure is
modeled using the temporal feature maps of a 3-D CNN, while
a temporal attention mechanism is used to combine information
across the entire video. For each generated word, the model can
focus on different temporal regions in the video. For simplicity, we
highlight only the region having the maximum attention above.",0.75,0.3311,0.5405,494,202,2.446,6cf37da62412dc5719a9a6e4af779963,images/2015/arxiv_0000114.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000115,Figure 115,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 3. Illustration of the proposed temporal attention mecha-
nism in the LSTM decoder",A diagram of the process of generating.,"A detailed layout showing the location of the camera ' s lens and the camera sensor, and other features.","A comprehensive technical explanation of the application of different types of motion on an airplane in flight, from left to right, to the camera, and the control system, with the pilot ' s position, the camera ' s view.","Figure 3. Illustration of the proposed temporal attention mecha-
nism in the LSTM decoder",0.7814,0.2393,0.5103,438,258,1.698,cd87e8775d20a39909465b3bc5441b7a,images/2015/arxiv_0000115.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000116,Figure 116,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure
2.
Illustration
of
the
spatio-
temporal convolutional neural network
(3-D CNN). This network is trained for
activity recognition. Then, only the con-
volutional layers are involved when gen-
erating video descriptions.",A diagram of the flow of data from two.,"A detailed layout showing the three different types of the camera and its components, including the camera lens.","A comprehensive technical explanation of the new video - based network for the virtual world, with a description of each component and its components, and an overview of the potential features, p2 - p3 - d - dn.","Figure
2.
Illustration
of
the
spatio-
temporal convolutional neural network
(3-D CNN). This network is trained for
activity recognition. Then, only the con-
volutional layers are involved when gen-
erating video descriptions.",0.75,0.2317,0.4909,960,396,2.424,2553480dbe633e4572953ea413eace59,images/2015/arxiv_0000116.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000117,Figure 117,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 4. Four sample videos and their corresponding generated and ground-truth descriptions from Youtube2Text (Left Column) and
DVS (Right Column). The bar plot under each frame corresponds to the attention weight αt
i for the frame when the corresponding word
(color-coded) was generated. From the top left panel, we can see that when the word “road” is about to be generated, the model focuses
highly on the third frame where the road is clearly visible. Similarly, on the bottom left panel, we can see that the model attends to the
second frame when it was about to generate the word “Someone”. The bottom row includes alternate descriptions generated by the other
model variations.",A diagram of a video that shows.,A detailed layout showing the sequence of a movie with multiple lines of color and images of people in different stages.,"A comprehensive technical explanation of a video production process for a film, with multiple stages and levels of performance and content creation in one video screener ' s life story, and the same time, each image, each.","Figure 4. Four sample videos and their corresponding generated and ground-truth descriptions from Youtube2Text (Left Column) and
DVS (Right Column). The bar plot under each frame corresponds to the attention weight αt
i for the frame when the corresponding word
(color-coded) was generated. From the top left panel, we can see that when the word “road” is about to be generated, the model focuses
highly on the third frame where the road is clearly visible. Similarly, on the bottom left panel, we can see that the model attends to the
second frame when it was about to generate the word “Someone”. The bottom row includes alternate descriptions generated by the other
model variations.",0.75,0.3366,0.5433,1224,606,2.02,ac6308d07633cf518470e44517161d81,images/2015/arxiv_0000117.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000118,Figure 118,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 6. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",A diagram of the different types of a.,A detailed layout showing the number of rows in a computer screen with blue and white lines on them.,"A comprehensive technical explanation of the different types of the same pattern and the different sizes of them, including the size, width, and length of each piece of each of the line, in a single file file or series,.","Figure 6. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",0.7071,0.3442,0.5256,1629,778,2.094,dc78d718b830a55ed39b961eed170715,images/2015/arxiv_0000118.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000119,Figure 119,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 6. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",A diagram of the number of lines that.,A detailed layout showing the sequence of a single - line pattern for the same pattern as shown in the following.,"A comprehensive technical explanation of the lvsmr - 1 and its applications in the process of using lasers to measure the length of the line segments of the object on the line of the same plane or the object,.","Figure 6. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",0.75,0.2617,0.5059,1224,1345,0.91,cd909d482fc9d3c3dc72efd4bd01d979,images/2015/arxiv_0000119.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000120,Figure 120,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 7. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. It is doing a better job at guessing the object being
chopped , compared with Figure 6.",A diagram of a computer screen with.,A detailed layout showing the rows of blue and white squares in a grided pattern with different colors.,"A comprehensive technical explanation of the first - generation pattern for the first generation of the internet network, using the same data to create a grid of squares and pixels in each one image, as well as well, as shown.","Figure 7. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. It is doing a better job at guessing the object being
chopped , compared with Figure 6.",0.7179,0.3302,0.5241,1627,791,2.057,a87af97bf1b95419eb78e36b7a7d9872,images/2015/arxiv_0000120.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000121,Figure 121,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 8. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",A diagram of the different types of.,A detailed layout showing the number and type of lines used to create a pixel pattern of squares and rectangles.,"A comprehensive technical explanation of the sound of a giraffe ' s ear and neck, including the sound waves and sound waves of its ears and earphones, and headphones, including a gir and earbuds representing all details, specifications, and.","Figure 8. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",0.75,0.3137,0.5318,1627,774,2.102,e014efeb966b75c34b82af5cc9fe88b7,images/2015/arxiv_0000121.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000122,Figure 122,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 7. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. It is doing a better job at guessing the object being
chopped , compared with Figure 6.",A diagram of the number of rows of.,A detailed layout showing the rows of blue and white lines that are used to create a pattern for a wall.,"A comprehensive technical explanation of the dna structure and function of human cell membranes, including the cellular structure and the cells in the cell wall, the cell membrane, and the cell walls, and its structures, and tissues,.","Figure 7. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. It is doing a better job at guessing the object being
chopped , compared with Figure 6.",0.7214,0.2847,0.503,1224,754,1.623,f9aa1d8b5a2026a4b922f045fa586d41,images/2015/arxiv_0000122.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000123,Figure 123,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 8. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",A diagram of a number of different.,"A detailed layout showing a number of lines of different sizes and colors, each with a different pattern.","A comprehensive technical explanation of the radio signal and the signal signal for the radio, including the signal number and the corresponding signals for each signal in the signal band and the receiver - on the radio speaker - line signal -.","Figure 8. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude of
α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",0.725,0.2899,0.5074,1224,521,2.349,26ee8b88dc084f618ac24ea7698b521f,images/2015/arxiv_0000123.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000124,Figure 124,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 9. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. The use of additional motion features offers more
faithful description of the action than the one without (“running” v.s. “walking” in Figure 8).",A diagram of the various types of.,"A detailed layout showing the number of different types of a sheep in the field, with a description of each.","A comprehensive technical explanation of the different levels of a data visual system for the data visual network of the united states, and canada, as well as described by the data source of the data from the data processing and visual systems.","Figure 9. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. The use of additional motion features offers more
faithful description of the action than the one without (“running” v.s. “walking” in Figure 8).",0.7357,0.3162,0.526,1627,775,2.099,d7efb404b2b032bf838966af6e9132a1,images/2015/arxiv_0000124.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000125,Figure 125,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 10. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude
of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",A diagram of a fence with different.,"A detailed layout showing the various patterns of the grass in the field, and the size of each of the trees.","A comprehensive technical explanation of a computer keyboard keyboard keyboard keys and numbers on a computer screen with lines in blue on the screen, and a keyboard keyboard on the keyboard keyboard with a number of a mouse in the keyboard,.","Figure 10. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude
of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",0.7357,0.2552,0.4955,1626,789,2.061,76130e1ad5941e5bd836c42ee1303159,images/2015/arxiv_0000125.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000126,Figure 126,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 9. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. The use of additional motion features offers more
faithful description of the action than the one without (“running” v.s. “walking” in Figure 8).",A diagram of the number of different.,"A detailed layout showing the number and type of lines for each type of instrument in a band, with an additional band.","A comprehensive technical explanation of the data visual of a computer system, including a clock and a line graph on it ' s screen, with a text box below that reads ' s name and a picture of a blue line of the same.","Figure 9. Model type: Basic + Local + Global. Model shifts its attention across frames when generating the caption. The bar char shows
the magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. The use of additional motion features offers more
faithful description of the action than the one without (“running” v.s. “walking” in Figure 8).",0.75,0.2595,0.5048,1224,660,1.855,e4c28ffc3fed403baafb093e0c7b8105,images/2015/arxiv_0000126.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000127,Figure 127,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 10. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude
of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",A diagram of a fence with blue and.,"A detailed layout showing the number of plants in the garden and the size of the fenced area that shows key features, attributes, and data points in.","A comprehensive technical explanation of the process for the first batch of the giraffes, including the following sequence of the sequence of each of the two girafeds, and the gir, orth representing all details, specifications, and configurations.","Figure 10. Model type: Basic + Global. Model shifts its attention across frames to generate a caption. The bar char shows the magitude
of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude.",0.81,0.2831,0.5466,1224,587,2.085,f3552a41be441a65b6d0d0008d68bafd,images/2015/arxiv_0000127.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000128,Figure 128,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 11. Model type: Basic + Local + Global. Model shifts its attention across frames to generate a caption. The bar char shows the
magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. 3DConvatt generates a more faithful description with a
much richer content than Figure 10. It even learns to generate a rare work “teasing”.",A diagram of the number of plants in a.,A detailed layout showing the number of times and locations of a train as it passes through the city.,"A comprehensive technical explanation of the data table for the data visual system, showing the data and the data information of each device in the figure below, as well as well, as described by the data source for the picture,.","Figure 11. Model type: Basic + Local + Global. Model shifts its attention across frames to generate a caption. The bar char shows the
magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. 3DConvatt generates a more faithful description with a
much richer content than Figure 10. It even learns to generate a rare work “teasing”.",0.7107,0.2511,0.4809,1628,794,2.05,398508bea82888cbcea5840a1ec65f77,images/2015/arxiv_0000128.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000129,Figure 129,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 11. Model type: Basic + Local + Global. Model shifts its attention across frames to generate a caption. The bar char shows the
magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. 3DConvatt generates a more faithful description with a
much richer content than Figure 10. It even learns to generate a rare work “teasing”.",A diagram of a field with a line of.,"A detailed layout showing the field of grass and blue squares on the left side of the screen, and the green and white squares on top of the right.","A comprehensive technical explanation of the cellular lattice in the united states of america, from the u s department of defense, to the u n s army and us military service corps of engineers, and navy, and the navy,.","Figure 11. Model type: Basic + Local + Global. Model shifts its attention across frames to generate a caption. The bar char shows the
magitude of α, sum to 1 row-wise, the higher the bar, the bigger the magnitude. 3DConvatt generates a more faithful description with a
much richer content than Figure 10. It even learns to generate a rare work “teasing”.",0.75,0.2833,0.5167,1224,985,1.243,e821362616178c3b7e36bba3a1164f39,images/2015/arxiv_0000129.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000130,Figure 130,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 12. Model type: Basic + Global. The model tends to produce a smooth distribution in α row-wise, due to the uniformity of the
scene with a slowly changing continuous shot.",A diagram of the number of columns in.,A detailed layout showing the number of blue bars in each row and the number in each line of the same.,"A comprehensive technical explanation of the sound waves in the sound wave chart for the soundwaves and the sound source for the sounds of the instrument, including the speaker, the sound system, the speakers, and the speaker.","Figure 12. Model type: Basic + Global. The model tends to produce a smooth distribution in α row-wise, due to the uniformity of the
scene with a slowly changing continuous shot.",0.7143,0.3355,0.5249,843,724,1.164,36b5574569629a8f2c11058ba6b10782,images/2015/arxiv_0000130.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000131,Figure 131,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 12. Model type: Basic + Global. The model tends to produce a smooth distribution in α row-wise, due to the uniformity of the
scene with a slowly changing continuous shot.",A diagram of the number of different.,"A detailed layout showing the number of lines of blue in each column for different types of data, including.","A comprehensive technical explanation of the development and operation of the d - 3e gene and simulation in pkt ' s report, with illustrations from the following slides, and illustrations of the results from the first two authors.","Figure 12. Model type: Basic + Global. The model tends to produce a smooth distribution in α row-wise, due to the uniformity of the
scene with a slowly changing continuous shot.",0.7357,0.2946,0.5151,1224,1116,1.097,b05c45ddf86e8a4c026741cba916f94e,images/2015/arxiv_0000131.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000132,Figure 132,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 13. Model type: Basic + Local + Global. The model learns a smooth α on the slowly changing scene. It captures a different action
from Basic + Global in Figure 12.",A diagram of a number of blue lines in.,"A detailed layout showing the number of columns in the data field of a computer screen, with the corresponding numbers.","A comprehensive technical explanation of the structure of a computer system and its characteristics, including the information and functions of the system and the design of it in the process and its application of its uses, pngwwhics.","Figure 13. Model type: Basic + Local + Global. The model learns a smooth α on the slowly changing scene. It captures a different action
from Basic + Global in Figure 12.",0.75,0.3208,0.5354,842,746,1.129,1459e03de614605cb4e63987ca28494d,images/2015/arxiv_0000132.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000133,Figure 133,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 13. Model type: Basic + Local + Global. The model learns a smooth α on the slowly changing scene. It captures a different action
from Basic + Global in Figure 12.",A diagram of a number of blue bars.,"A detailed layout showing the number of lines of blue in each row, and the number on the right that shows key features, attributes, and data points.","A comprehensive technical explanation of the number of lines in the data table, using the following numbers, and the corresponding numbers, as well as the time of each line of each element, of the data, for each time,.","Figure 13. Model type: Basic + Local + Global. The model learns a smooth α on the slowly changing scene. It captures a different action
from Basic + Global in Figure 12.",0.81,0.2998,0.5549,1224,1177,1.04,79851b251e777c4c3f86d68f993f7c8b,images/2015/arxiv_0000133.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000134,Figure 134,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,Figure 14. Model type: Basic + Global. α also reflects the sudden transition between two shots.,A diagram of the various types of.,"A detailed layout showing the number of rows of blue and white lines in the same row, each with different.","A comprehensive technical explanation of the human body ' s blood cycle - - - and - its structure, including the blood cycle and the blood flow and the body of the blood system of the water source of the body,.",Figure 14. Model type: Basic + Global. α also reflects the sudden transition between two shots.,0.7286,0.3322,0.5304,846,783,1.08,74ff2c5c80c4f3fc0c58652065e61469,images/2015/arxiv_0000134.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000135,Figure 135,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,Figure 14. Model type: Basic + Global. α also reflects the sudden transition between two shots.,A diagram of a computer screen showing.,"A detailed layout showing the blue and white patterns of a dna strip and a pattern of blue squares that shows key features, attributes, and data.","A comprehensive technical explanation of the structure of a gene, including the dna and its structure and function, as well as the number of cells and the size of the size and type of the individual dnas, of the gene.",Figure 14. Model type: Basic + Global. α also reflects the sudden transition between two shots.,0.81,0.3055,0.5577,772,1209,0.639,eeb4a2ab5ddd49ca987977144dea68a0,images/2015/arxiv_0000135.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000136,Figure 136,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 15. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 14, attempting
to incoporate character-level interaction inside the first part of the scene.",A diagram of the number and type of.,A detailed layout showing the number of different types of blue squares in each row and a line of smaller blue.,"A comprehensive technical explanation of the human body ' s genetic structure and function in the body and the brain, including the structure of the body, the human, and the body ' dysmeting, the nervous,.","Figure 15. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 14, attempting
to incoporate character-level interaction inside the first part of the scene.",0.7464,0.3092,0.5278,845,796,1.062,889f780d79315d5ed4cf89f27c76e124,images/2015/arxiv_0000136.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000137,Figure 137,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 15. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 14, attempting
to incoporate character-level interaction inside the first part of the scene.",A diagram of a large number of blue.,"A detailed layout showing the number of different types of blue lines in a grid pattern, in a white background.","A comprehensive technical explanation of the first step in the data visual tool for the web page on a computer screen, including data visual tools and a graph bar chart of a set of data visual data visual bar chart on top.","Figure 15. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 14, attempting
to incoporate character-level interaction inside the first part of the scene.",0.7464,0.3313,0.5388,1224,1205,1.016,7740f413e926e5af8c6a5987d9951833,images/2015/arxiv_0000137.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000138,Figure 138,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 16. Model type: Basic + Global. The model seems to focus on the second shot of the scene at the beginning, yet the part of the
generated caption “out of the car” distributes a decent amount of its attention on the first scene as well. This may due to the fact that the
memory of decoding LSTM already contains the information of almost the entire scene (two shots).",A diagram of a cell phone with a long.,"A detailed layout showing the number of rows of blue lines for each of the lines in this picture that shows key features, attributes, and data.","A comprehensive technical explanation of the pattern of the first row of rows in an array of blue lines on white paper, with a black border, and a white background that includes a red line, and black, black, and blue.","Figure 16. Model type: Basic + Global. The model seems to focus on the second shot of the scene at the beginning, yet the part of the
generated caption “out of the car” distributes a decent amount of its attention on the first scene as well. This may due to the fact that the
memory of decoding LSTM already contains the information of almost the entire scene (two shots).",0.81,0.3146,0.5623,852,774,1.101,957cc57fd4608f9182300582df7aee2c,images/2015/arxiv_0000138.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000139,Figure 139,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 16. Model type: Basic + Global. The model seems to focus on the second shot of the scene at the beginning, yet the part of the
generated caption “out of the car” distributes a decent amount of its attention on the first scene as well. This may due to the fact that the
memory of decoding LSTM already contains the information of almost the entire scene (two shots).",A diagram of a number of rows of blue.,A detailed layout showing the rows of blue lines in a white background with a black border and a red dot.,"A comprehensive technical explanation of the first generation of the microconvergats for the new generation of microconversity in the u s v i - v r - v system, as well, and the u - v representing all details, specifications, and configurations of.","Figure 16. Model type: Basic + Global. The model seems to focus on the second shot of the scene at the beginning, yet the part of the
generated caption “out of the car” distributes a decent amount of its attention on the first scene as well. This may due to the fact that the
memory of decoding LSTM already contains the information of almost the entire scene (two shots).",0.725,0.2709,0.498,1224,1212,1.01,f6e75f4e71c9a57829b5cd5f20b861bc,images/2015/arxiv_0000139.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000140,Figure 140,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 17. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 16. The model
focuses on the car in the second shot when generating “sit”, “back seat”. When generating two “SOMEONE”, it divides its attentio among
two shots.",A diagram of a set of blue bars on a.,A detailed layout showing the different types of lines used to create a grid pattern for each of the three rows.,"A comprehensive technical explanation of the mean of a data visual representation in a computer screen shot from the web site, showing a pattern of blue lines with dots on them in rows and verticals and vertical rows of varying sizes.","Figure 17. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 16. The model
focuses on the car in the second shot when generating “sit”, “back seat”. When generating two “SOMEONE”, it divides its attentio among
two shots.",0.75,0.2953,0.5227,843,779,1.082,cf191bc846dc5bc83fefa6a21f0a257f,images/2015/arxiv_0000140.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000141,Figure 141,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 17. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 16. The model
focuses on the car in the second shot when generating “sit”, “back seat”. When generating two “SOMEONE”, it divides its attentio among
two shots.",A diagram of the same type of lines in.,A detailed layout showing the number of rows of blue and white lines on a white background with a black border.,"A comprehensive technical explanation of the first generation of the hypothermic gene structure and its origins, including the hystothermyal and hypodiume genes and the hysphaic gene representing all details, specifications, and configurations of.","Figure 17. Model type: Basic + Local + Global. The learned model generates a more sophiscated description than Figure 16. The model
focuses on the car in the second shot when generating “sit”, “back seat”. When generating two “SOMEONE”, it divides its attentio among
two shots.",0.7464,0.2929,0.5196,1224,1217,1.006,eff363f4a5791676b1132ab5d9904164,images/2015/arxiv_0000141.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000142,Figure 142,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,Figure 18. Model type: Basic + Global. The description is argubly not very accurate,A diagram of a blue and white pattern.,"A detailed layout showing the pattern of blue bars and dots on white paper, as well as a line drawing.","A comprehensive technical explanation of the linear pattern of a square - shaped object, including the number of squares and the number for each one side of the figure in the space in the figure, and the image, with a rectangle.",Figure 18. Model type: Basic + Global. The description is argubly not very accurate,0.7143,0.2872,0.5008,849,776,1.094,912464640b7d55edf4fe0d8dc00b509b,images/2015/arxiv_0000142.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000143,Figure 143,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,Figure 18. Model type: Basic + Global. The description is argubly not very accurate,A diagram of the blue lines are.,"A detailed layout showing the dna of a human cell phone phone, with blue and white lines on it that shows key features, attributes, and data points.","A comprehensive technical explanation of the dna of the human body, from the earliest to present in the modern day, and beyond the present time period of life, the evolution of the study of the cereuterment period of the genetic.",Figure 18. Model type: Basic + Global. The description is argubly not very accurate,0.81,0.2676,0.5388,685,1236,0.554,db80ecb6a47a5f9169e4ee21ffda8fa3,images/2015/arxiv_0000143.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000144,Figure 144,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 19. Model type: Basic + Local + Global. With the help of additional features, the model successfully describes the cell phone and
the room, a much faithful description than Figure 18",A diagram of the different types of.,"A detailed layout showing the different sizes and types of cell phones and their functions, including the number of cellphones.","A comprehensive technical explanation of the blue lines and the different types of cells in the cell phone screen are shown in this diagram, with the cellphone number of the cell phones in the blue is the cell of the numbers.","Figure 19. Model type: Basic + Local + Global. With the help of additional features, the model successfully describes the cell phone and
the room, a much faithful description than Figure 18",0.75,0.2496,0.4998,845,786,1.075,111ac28932b1cead9fe5bb81999a2da7,images/2015/arxiv_0000144.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000145,Figure 145,scientific_figure,Describing Videos by Exploiting Temporal Structure,arXiv Research Authors,1502.08029v5,stat,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). Describing Videos by Exploiting Temporal Structure. arXiv:1502.08029v5,"Figure 19. Model type: Basic + Local + Global. With the help of additional features, the model successfully describes the cell phone and
the room, a much faithful description than Figure 18",A diagram of a large number of blue.,A detailed layout showing the number of rows and the names of each type of pattern in blue on white.,"A comprehensive technical explanation of the structure and function of a computer system, including its components and functions, for each user of the device, in the system ' s design and operating environment ' s environment ' d ' s own.","Figure 19. Model type: Basic + Local + Global. With the help of additional features, the model successfully describes the cell phone and
the room, a much faithful description than Figure 18",0.7071,0.3236,0.5153,1224,1230,0.995,993d6093fd66a7520b32bc5b12af65c9,images/2015/arxiv_0000145.png,https://arxiv.org/pdf/1502.08029v5.pdf
arxiv_0000146,Figure 146,scientific_figure,A Unified Deep Neural Network for Speaker and Language Recognition,arXiv Research Authors,1504.00923v1,cs,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). A Unified Deep Neural Network for Speaker and Language Recognition. arXiv:1504.00923v1,Fig. 2. Example DNN architecture,A diagram of the different paths of a.,"A detailed layout showing the different paths of the dots in the diagram, and the corresponding dots on each side.","A comprehensive technical explanation of the dot - check algorithm and its applications, part 2 / 3, p 1 / 2 / 4, p 3 / 4 / 6 / 7 / 8 / 8 p 9 / 12 / 2 representing all details, specifications, and configurations of the system components in full.",Fig. 2. Example DNN architecture,0.81,0.2791,0.5446,599,354,1.692,153427a5d9e0b6801d1380b43b53dc91,images/2015/arxiv_0000146.png,https://arxiv.org/pdf/1504.00923v1.pdf
arxiv_0000147,Figure 147,scientific_figure,A Unified Deep Neural Network for Speaker and Language Recognition,arXiv Research Authors,1504.00923v1,cs,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). A Unified Deep Neural Network for Speaker and Language Recognition. arXiv:1504.00923v1,"Fig. 1. Simplified block diagram of i-vector extraction and
scoring.",A diagram of a block with multiple.,"A detailed layout showing the structure of an external memory processor and its components and functions, including the memory memory.","A comprehensive technical explanation of the super - vector function of a single - channel television receiver and receiver circuit diagram from figure 1, from the electronic signal architecture library of the library of sonyn, sony, volt, 1978.","Fig. 1. Simplified block diagram of i-vector extraction and
scoring.",0.75,0.2889,0.5194,697,170,4.1,aa035323cb2e08d228b9384813c9ee22,images/2015/arxiv_0000147.png,https://arxiv.org/pdf/1504.00923v1.pdf
arxiv_0000148,Figure 148,scientific_figure,A Unified Deep Neural Network for Speaker and Language Recognition,arXiv Research Authors,1504.00923v1,cs,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). A Unified Deep Neural Network for Speaker and Language Recognition. arXiv:1504.00923v1,"Fig. 1. Simplified block diagram of i-vector extraction and
scoring.",A diagram of a block diagram with.,"A detailed layout showing the process of the proposed system for the computer system, including the processor and the processor.","A comprehensive technical explanation of the new computer system for the ibm system, including the intel processor and the intel i / o processor, from ibm - tw - series 1 - 3 0 and intel - 2 - t - 8 representing all details, specifications, and.","Fig. 1. Simplified block diagram of i-vector extraction and
scoring.",0.75,0.2495,0.4998,569,254,2.24,1d4931a147f24d7512d37164d63e2363,images/2015/arxiv_0000148.png,https://arxiv.org/pdf/1504.00923v1.pdf
arxiv_0000149,Figure 149,scientific_figure,A Unified Deep Neural Network for Speaker and Language Recognition,arXiv Research Authors,1504.00923v1,cs,2015,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2015). A Unified Deep Neural Network for Speaker and Language Recognition. arXiv:1504.00923v1,Fig. 2. Example DNN architecture,A diagram of a network with many dots.,A detailed layout showing the structure of the dot - checker and its corresponding positions and functions in a single - dimensional model.,"A comprehensive technical explanation of a networked data processor for the internet age system, with a number of different connections in each region and a total number of bits in the networked networked bits permuts of each one.",Fig. 2. Example DNN architecture,0.81,0.2844,0.5472,358,255,1.404,0fa7c1d768cde1c6e6d8eb29a077368a,images/2015/arxiv_0000149.png,https://arxiv.org/pdf/1504.00923v1.pdf
arxiv_0000150,Figure 150,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Root
I
saw
a
friend
today",A diagram of the components for an.,"A detailed layout showing different features for each of the three phases of the project, including the following project.","A comprehensive technical explanation of the web user ' s experience and features for the web application in this diagram, we see all three components of the frameworks in this image below in the following figure 2 / 3 / 4 /.","Root
I
saw
a
friend
today",0.81,0.2683,0.5392,556,204,2.725,d732e7b24f82f35116404b62df09e056,images/2016/arxiv_0000150.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000151,Figure 151,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Figure 1: Left: When predicting the dependency tree, some dependencies (e.g., the dashed edges) are easily
resolved, and there is less need for expressive features in making a prediction. Right: Our system diagram and
operating region. When relatively feature and inference costs are both not insignificant the policy must carefully
balance the overhead costs due to feedback from the predictor with feature acquisition costs.",A diagram of the steps in the process.,"A detailed layout showing the steps of a four - way cycle of wind direction from left to right that shows key features, attributes, and data points.","A comprehensive technical explanation of root, saw, and friend - to - stay together in the same direction of the tree and the wind is moving through the earth and the same path on the same plane as the same way,.","Figure 1: Left: When predicting the dependency tree, some dependencies (e.g., the dashed edges) are easily
resolved, and there is less need for expressive features in making a prediction. Right: Our system diagram and
operating region. When relatively feature and inference costs are both not insignificant the policy must carefully
balance the overhead costs due to feedback from the predictor with feature acquisition costs.",0.81,0.2812,0.5456,349,107,3.262,552c4065afa8ec141ee5e044bd519cfc,images/2016/arxiv_0000151.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000152,Figure 152,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Figure 3: The performance of our one-shot policy (in red) is compared to the uniform strategy (in
black) and policy of Weiss et. al. [25] for the OCR dataset. Although the policy with complex
features is more efficient with features, the simple feature policy has a lower total run-time in the low
budget region due to the overhead of additional inference. [18] to the structured prediction case. The myopic policy runs the structured predictor initially on
all cheap features, then looks at the total confidence of the classifier normalized by the sample size
(e.g. sentence length). If the confidence is below a threshold, it chooses to acquire expensive features
for all positions. Finally, we compare against the Q-learning method proposed by [25]. This method
requires global features for structures with varying size. From now on we will refer to features that
require access to more than one part as complex features and part level features as simple features.
In their case, they use confidence feedback from the structured predictor which induces additional
inference overhead for the policy. In addition to this, it is not straightforward to apply this approach
to do part by part feature selection on structures with varying sizes. We adopt Structured-SVM [19] to solve the policy learning problems for expected and anytime cases
defined in (4) and (8), respectively. For the structure of the policy π we use a graph with no edges
due to its simplicity. In this form, the policy learning problem can be written as a sample weighted
SVM. We discuss the details in the appendix due to space constraints. We show in the following that complex features indeed benefit the policy, but simple features perform
better for cases where the inference time and feature costs are comparable and the additional overhead
is unwanted. Finally, we show that part by part selection outperforms global selection. Optical Character Recognition We tested our algorithm on a sequence-label problem, the OCR
dataset [17] composed of 6,877 handwritten words, where each word is represented as a sequence of
16x8 binary letter images. We use a linear-chain Markov model, and similar to the setup in [23, 21],
use raw pixel values and HOG features with 3x3 cell size as our feature templates. We split the data
such that 90% percent is used for training and 10% is used for test.",A diagram of the curve in which a.,"A detailed layout showing the time of each race in the year, as well as the actual time of the race.","A comprehensive technical explanation of the actual time and average running times for the course in the course course, and the course time for the race, and course, the course, course, or event, course and course course.","Figure 3: The performance of our one-shot policy (in red) is compared to the uniform strategy (in
black) and policy of Weiss et. al. [25] for the OCR dataset. Although the policy with complex
features is more efficient with features, the simple feature policy has a lower total run-time in the low
budget region due to the overhead of additional inference. [18] to the structured prediction case. The myopic policy runs the structured predictor initially on
all cheap features, then looks at the total confidence of the classifier normalized by the sample size
(e.g. sentence length). If the confidence is below a threshold, it chooses to acquire expensive features
for all positions. Finally, we compare against the Q-learning method proposed by [25]. This method
requires global features for structures with varying size. From now on we will refer to features that
require access to more than one part as complex features and part level features as simple features.
In their case, they use confidence feedback from the structured predictor which induces additional
inference overhead for the policy. In addition to this, it is not straightforward to apply this approach
to do part by part feature selection on structures with varying sizes. We adopt Structured-SVM [19] to solve the policy learning problems for expected and anytime cases
defined in (4) and (8), respectively. For the structure of the policy π we use a graph with no edges
due to its simplicity. In this form, the policy learning problem can be written as a sample weighted
SVM. We discuss the details in the appendix due to space constraints. We show in the following that complex features indeed benefit the policy, but simple features perform
better for cases where the inference time and feature costs are comparable and the additional overhead
is unwanted. Finally, we show that part by part selection outperforms global selection. Optical Character Recognition We tested our algorithm on a sequence-label problem, the OCR
dataset [17] composed of 6,877 handwritten words, where each word is represented as a sequence of
16x8 binary letter images. We use a linear-chain Markov model, and similar to the setup in [23, 21],
use raw pixel values and HOG features with 3x3 cell size as our feature templates. We split the data
such that 90% percent is used for training and 10% is used for test.",0.7071,0.2576,0.4823,560,420,1.333,3c9ad2f67eb1ef48fd9c6098394310f7,images/2016/arxiv_0000152.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000153,Figure 153,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Figure 2: An example word from the OCR
test dataset is shown. Note that the word is
initially incorrectly identified due to degra-
dation in letters ""u"" and ""n"". The letter clas-
sification accuracy increases after the policy
acquires the HOG features at strategic posi-
tions.",A diagram of the number of letters.,A detailed layout showing the pattern of the first layer of the image and the second layer of its image.,"A comprehensive technical explanation of the data visual interface for the x - ray camera ' s view of the screen and its size, and the size, of the time, and location, the time and the object representing all details, specifications, and.","Figure 2: An example word from the OCR
test dataset is shown. Note that the word is
initially incorrectly identified due to degra-
dation in letters ""u"" and ""n"". The letter clas-
sification accuracy increases after the policy
acquires the HOG features at strategic posi-
tions.",0.7814,0.2778,0.5296,1374,470,2.923,a426f9a4721cb6e891b307b7b52c1796,images/2016/arxiv_0000153.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000154,Figure 154,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Figure 2: An example word from the OCR
test dataset is shown. Note that the word is
initially incorrectly identified due to degra-
dation in letters ""u"" and ""n"". The letter clas-
sification accuracy increases after the policy
acquires the HOG features at strategic posi-
tions.",A diagram of the sound and language of.,A detailed layout showing the keyboard of a computer with a black and white image of a keyboard with a yellow box.,"A comprehensive technical explanation of the letter ' s name and its meaning in the alphabet, in the form of an x - ray image on the keyboard keyboard keyboard screen, with a green arrow pointing toward the word of the letters.","Figure 2: An example word from the OCR
test dataset is shown. Note that the word is
initially incorrectly identified due to degra-
dation in letters ""u"" and ""n"". The letter clas-
sification accuracy increases after the policy
acquires the HOG features at strategic posi-
tions.",0.75,0.2569,0.5034,400,300,1.333,b45db90ee7f7aaf9dfc7396f7324f575,images/2016/arxiv_0000154.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000155,Figure 155,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Figure 4: Left: Performance of various adaptive policies for varying budget levels (dependency tree accuracy
vs. total execution time), is compared to a uniform strategy on word and sentence level, and myopic policy
for the 23 section of PTB dataset. Right: Distribution of parse-tree depth for words that use cheap (green) or
expensive features (orange) for anytime policy. Time increases from left to right. Each group of columns show
the distribution of depths from 0(root) to 7. The policy is concentrated on acquiring features for lower depth
words. A sentence example also shows this effect. It is easy to identify parents of the adjectives and determiner.
However, additional features(orange) are required for the root(verb), subject and object.",A diagram of a curve with different.,"A detailed layout showing the relationship of a time - and - date policy with a timezone of 12, 000.","A comprehensive technical explanation of the time - to - date and time - period model for the new system, including the u s and u s policy policy policy, from the current policy of the same countries, and the end,.","Figure 4: Left: Performance of various adaptive policies for varying budget levels (dependency tree accuracy
vs. total execution time), is compared to a uniform strategy on word and sentence level, and myopic policy
for the 23 section of PTB dataset. Right: Distribution of parse-tree depth for words that use cheap (green) or
expensive features (orange) for anytime policy. Time increases from left to right. Each group of columns show
the distribution of depths from 0(root) to 7. The policy is concentrated on acquiring features for lower depth
words. A sentence example also shows this effect. It is easy to identify parents of the adjectives and determiner.
However, additional features(orange) are required for the root(verb), subject and object.",0.7071,0.2415,0.4743,714,566,1.261,ae3d7e148ce46468905fc0c5820c6f68,images/2016/arxiv_0000155.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000156,Figure 156,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Fig. 4 shows the test performance (unlabeled attachment accuracy) along with inference time. We
see that all one-shot policies perform similarly, losing negligible accuracy when using half of the
available expensive features. When we apply the length dictionary filtering heuristic in [7, 13], our
parser achieves 89.7% UAS on PTB section 23 with overall running time merely 7.5 seconds (I/O
excluded, 10s with I/O) and obtains 2.9X total speed-up with losing only 1% UAS comparing to the
baseline. 4 This significant speed-up over an efficient implementation is remarkable. 5 Although marginal, one-shot policy with greedy trajectory has the strongest performance in low
budget regions. This is because the greedy trajectory search has better granularity than parsimonious
search in choosing positions that decrease the loss early on. The anytime policy is below one-shot
policy for all budget levels. As discussed in 2.2, the anytime policy is more constrained in that it has
to achieve a fixed budget for all examples. The naive myopic policy performs worse than uniform
since it has to run inference on samples with low confidence two times, adding approximately 4.5
seconds of extra time for the full test dataset. We then explore the effect of importance weights for
the greedy policy. We notice a small improvement. We hypothesize that this is due to the policy
functional complexity being a limiting factor. We also conduct ablative studies to better understand the policy behavior. Fig. 4 shows the distribution
of depth for the words that use expensive and cheap features in the ground truth dependency tree. We
expect investing more time on the low-depth words (root in the extreme) to yield higher accuracy
gains. We observe this phenomenon empirically, as the policy concentrates on extracting features
close to the root.",A diagram of a tree with several.,A detailed layout showing the tree of the two classes in which the following classes are represented in the diagram.,"A comprehensive technical explanation of the root tree for the web application for a computer system, including a networked environment and a basic application for the internet system, as well known source, as described by a web, as source.","Fig. 4 shows the test performance (unlabeled attachment accuracy) along with inference time. We
see that all one-shot policies perform similarly, losing negligible accuracy when using half of the
available expensive features. When we apply the length dictionary filtering heuristic in [7, 13], our
parser achieves 89.7% UAS on PTB section 23 with overall running time merely 7.5 seconds (I/O
excluded, 10s with I/O) and obtains 2.9X total speed-up with losing only 1% UAS comparing to the
baseline. 4 This significant speed-up over an efficient implementation is remarkable. 5 Although marginal, one-shot policy with greedy trajectory has the strongest performance in low
budget regions. This is because the greedy trajectory search has better granularity than parsimonious
search in choosing positions that decrease the loss early on. The anytime policy is below one-shot
policy for all budget levels. As discussed in 2.2, the anytime policy is more constrained in that it has
to achieve a fixed budget for all examples. The naive myopic policy performs worse than uniform
since it has to run inference on samples with low confidence two times, adding approximately 4.5
seconds of extra time for the full test dataset. We then explore the effect of importance weights for
the greedy policy. We notice a small improvement. We hypothesize that this is due to the policy
functional complexity being a limiting factor. We also conduct ablative studies to better understand the policy behavior. Fig. 4 shows the distribution
of depth for the words that use expensive and cheap features in the ground truth dependency tree. We
expect investing more time on the low-depth words (root in the extreme) to yield higher accuracy
gains. We observe this phenomenon empirically, as the policy concentrates on extracting features
close to the root.",0.81,0.2934,0.5517,1820,1056,1.723,d97297f5d5dcf95ec7044adf879c739c,images/2016/arxiv_0000156.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000157,Figure 157,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Figure 4: Left: Performance of various adaptive policies for varying budget levels (dependency tree accuracy
vs. total execution time), is compared to a uniform strategy on word and sentence level, and myopic policy
for the 23 section of PTB dataset. Right: Distribution of parse-tree depth for words that use cheap (green) or
expensive features (orange) for anytime policy. Time increases from left to right. Each group of columns show
the distribution of depths from 0(root) to 7. The policy is concentrated on acquiring features for lower depth
words. A sentence example also shows this effect. It is easy to identify parents of the adjectives and determiner.
However, additional features(orange) are required for the root(verb), subject and object.",A diagram of the number of steps in a.,"A detailed layout showing the steps in the sequence of the process, including the number and step steps.","A comprehensive technical explanation of the average time of a team of players to play the game from each team on the field, including the start and finish line up stages, the next step one game of the next match, in the next.","Figure 4: Left: Performance of various adaptive policies for varying budget levels (dependency tree accuracy
vs. total execution time), is compared to a uniform strategy on word and sentence level, and myopic policy
for the 23 section of PTB dataset. Right: Distribution of parse-tree depth for words that use cheap (green) or
expensive features (orange) for anytime policy. Time increases from left to right. Each group of columns show
the distribution of depths from 0(root) to 7. The policy is concentrated on acquiring features for lower depth
words. A sentence example also shows this effect. It is easy to identify parents of the adjectives and determiner.
However, additional features(orange) are required for the root(verb), subject and object.",0.7214,0.301,0.5112,2086,974,2.142,c1997ca6c1f203db37b157974d8b8779,images/2016/arxiv_0000157.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000158,Figure 158,scientific_figure,Resource Constrained Structured Prediction,arXiv Research Authors,1602.08761v2,stat,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Resource Constrained Structured Prediction. arXiv:1602.08761v2,"Figure 4: Left: Performance of various adaptive policies for varying budget levels (dependency tree accuracy
vs. total execution time), is compared to a uniform strategy on word and sentence level, and myopic policy
for the 23 section of PTB dataset. Right: Distribution of parse-tree depth for words that use cheap (green) or
expensive features (orange) for anytime policy. Time increases from left to right. Each group of columns show
the distribution of depths from 0(root) to 7. The policy is concentrated on acquiring features for lower depth
words. A sentence example also shows this effect. It is easy to identify parents of the adjectives and determiner.
However, additional features(orange) are required for the root(verb), subject and object.",A diagram of a group of graphs and a.,A detailed layout showing the various processes of the experiment and the results of the results in the experiment.,"A comprehensive technical explanation of the model for a new system of multiple systems, including a graph and a line plot of data, and an image of a graphing function that shows a wave, as well described by the same.","Figure 4: Left: Performance of various adaptive policies for varying budget levels (dependency tree accuracy
vs. total execution time), is compared to a uniform strategy on word and sentence level, and myopic policy
for the 23 section of PTB dataset. Right: Distribution of parse-tree depth for words that use cheap (green) or
expensive features (orange) for anytime policy. Time increases from left to right. Each group of columns show
the distribution of depths from 0(root) to 7. The policy is concentrated on acquiring features for lower depth
words. A sentence example also shows this effect. It is easy to identify parents of the adjectives and determiner.
However, additional features(orange) are required for the root(verb), subject and object.",0.81,0.2456,0.5278,1224,488,2.508,f612dbd5d21b3935d02070b59ea3f4fc,images/2016/arxiv_0000158.png,https://arxiv.org/pdf/1602.08761v2.pdf
arxiv_0000159,Figure 159,scientific_figure,Natural-Parameter Networks: A Class of Probabilistic Neural Networks,arXiv Research Authors,1611.00448v1,cs,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Natural-Parameter Networks: A Class of Probabilistic Neural Networks. arXiv:1611.00448v1,"Figure 2:
Classification accuracy
for different variance (uncertainty).
Note that ‘1’ in the x-axis means
a(L)
s
1T ∈[0, 0.04), ‘2’ means
a(L)
s
1T ∈[0.04, 0.08), etc.",A diagram of a graph with the number.,A detailed layout showing a bar chart of hyperparameters and the average and average data for each type.,"A comprehensive technical explanation of the optimal hypparametrics for the data - driven system, using data - generated hyperparameters, and data - visual processing techniques, including data - based on the same platform,.","Figure 2:
Classification accuracy
for different variance (uncertainty).
Note that ‘1’ in the x-axis means
a(L)
s
1T ∈[0, 0.04), ‘2’ means
a(L)
s
1T ∈[0.04, 0.08), etc.",0.7814,0.3438,0.5626,224,157,1.427,ba815acb4ffc8be4cde195bd54bd24b5,images/2016/arxiv_0000159.png,https://arxiv.org/pdf/1611.00448v1.pdf
arxiv_0000160,Figure 160,scientific_figure,Natural-Parameter Networks: A Class of Probabilistic Neural Networks,arXiv Research Authors,1611.00448v1,cs,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Natural-Parameter Networks: A Class of Probabilistic Neural Networks. arXiv:1611.00448v1,"Figure 3: Reconstruction error
and estimated uncertainty for
each data point in Citeulike-a.",A diagram of the distribution of the.,"A detailed layout showing the differences between the left and right side of the ploter ' s plot that shows key features, attributes, and data.","A comprehensive technical explanation of the difference of the standard and standard vf parameters in human blood flow, from the study of the study on the human body ' s dys - related to - t - vy - t.","Figure 3: Reconstruction error
and estimated uncertainty for
each data point in Citeulike-a.",0.87,0.2856,0.5778,240,169,1.42,4c2ebf2786abbc1431d9c55fcb10151e,images/2016/arxiv_0000160.png,https://arxiv.org/pdf/1611.00448v1.pdf
arxiv_0000161,Figure 161,scientific_figure,Natural-Parameter Networks: A Class of Probabilistic Neural Networks,arXiv Research Authors,1611.00448v1,cs,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Natural-Parameter Networks: A Class of Probabilistic Neural Networks. arXiv:1611.00448v1,"Figure 4: Activation functions for the gamma distribution (left), the beta distribution (middle), and
the Rayleigh distribution (right).",A diagram of the timeframes for.,"A detailed layout showing the various different types of the data and graphs for each type of data, including data.","A comprehensive technical explanation of the mean and slope plot for the two types of high - speed trains, from the beginning to end of the 20th century, as described by the same time, in the following years of the train.","Figure 4: Activation functions for the gamma distribution (left), the beta distribution (middle), and
the Rayleigh distribution (right).",0.81,0.2285,0.5192,795,190,4.184,02b91a1cd41c56773fa12b4f4c69eb52,images/2016/arxiv_0000161.png,https://arxiv.org/pdf/1611.00448v1.pdf
arxiv_0000162,Figure 162,scientific_figure,Collaborative Recurrent Autoencoder: Recommend while Learning to Fill in the Blanks,arXiv Research Authors,1611.00454v1,cs,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Collaborative Recurrent Autoencoder: Recommend while Learning to Fill in the Blanks. arXiv:1611.00454v1,"Figure 1: On the left is the graphical model for an example CRAE where Tj = 2 for all j. To
prevent clutter, the hyperparameters for beta-pooling, all weights, biases, and links between ht and γ
are omitted. On the right is the graphical model for the degenerated CRAE. An example recurrent
autoencoder with Tj = 3 is shown. ‘⟨?⟩’ is the ⟨wildcard⟩and ‘$’ marks the end of a sentence. E′",A diagram of a cell phone that is.,A detailed layout showing the different stages of the model for a cell phone and a diagram of the cell phone.,"A comprehensive technical explanation of the proposed design for the first - generation hybrid - mode - battery system for the new - generation battery system, as described by the two separate models in the previous image from the model shown below,.","Figure 1: On the left is the graphical model for an example CRAE where Tj = 2 for all j. To
prevent clutter, the hyperparameters for beta-pooling, all weights, biases, and links between ht and γ
are omitted. On the right is the graphical model for the degenerated CRAE. An example recurrent
autoencoder with Tj = 3 is shown. ‘⟨?⟩’ is the ⟨wildcard⟩and ‘$’ marks the end of a sentence. E′",0.8593,0.2397,0.5495,755,213,3.545,60f37a71c7664badbe339800161a75f7,images/2016/arxiv_0000162.png,https://arxiv.org/pdf/1611.00454v1.pdf
arxiv_0000163,Figure 163,scientific_figure,Collaborative Recurrent Autoencoder: Recommend while Learning to Fill in the Blanks,arXiv Research Authors,1611.00454v1,cs,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Collaborative Recurrent Autoencoder: Recommend while Learning to Fill in the Blanks. arXiv:1611.00454v1,"Figure 2: Performance comparison of CRAE, CDL, CTR, DeepMusic, CMF, and SVDFeature based
on recall@M for datasets CiteULike and Netflix. P is varied from 1 to 5 in the first two figures.
We also use another evaluation metric, mean average precision (mAP), in the experiments. Exactly
the same as [10], the cutoff point is set at 500 for each user.",A diagram of a line graph with.,A detailed layout showing the different types of the various processes in the process of producing a new product.,"A comprehensive technical explanation of the three new phase - time data for the new model, with different data and results in the same set of stages of action, in the next phase, each phase, the following phase,.","Figure 2: Performance comparison of CRAE, CDL, CTR, DeepMusic, CMF, and SVDFeature based
on recall@M for datasets CiteULike and Netflix. P is varied from 1 to 5 in the first two figures.
We also use another evaluation metric, mean average precision (mAP), in the experiments. Exactly
the same as [10], the cutoff point is set at 500 for each user.",0.75,0.2081,0.479,810,170,4.765,d99e5d83e1146d48d7bc861303411bc7,images/2016/arxiv_0000163.png,https://arxiv.org/pdf/1611.00454v1.pdf
arxiv_0000164,Figure 164,scientific_figure,Collaborative Recurrent Autoencoder: Recommend while Learning to Fill in the Blanks,arXiv Research Authors,1611.00454v1,cs,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Collaborative Recurrent Autoencoder: Recommend while Learning to Fill in the Blanks. arXiv:1611.00454v1,Figure 5: The recall@M for different λv.,A diagram of the number of people in.,A detailed layout showing the linear and scatter function of the data visual system in the data processing system.,"A comprehensive technical explanation of the mean of the number of people in a long time period of birth and development, by age, from 2000 to 2007 - 2009 - 2006 - 201006, 0002, 000 - 004 representing all details, specifications, and configurations.",Figure 5: The recall@M for different λv.,0.75,0.26,0.505,328,230,1.426,197ee90636a27e424d6f5079a02ac0df,images/2016/arxiv_0000164.png,https://arxiv.org/pdf/1611.00454v1.pdf
arxiv_0000165,Figure 165,scientific_figure,Invariant Representations for Noisy Speech Recognition,arXiv Research Authors,1612.01928v1,cs,2016,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2016). Invariant Representations for Noisy Speech Recognition. arXiv:1612.01928v1,Figure 1: Model structure for invariant training and ASR results.,A diagram of the flow of a computer.,"A detailed layout showing the differences between the two types of noise and noise conditions in the system,.","A comprehensive technical explanation of the basic model for the noise condition and the effect of noise conditions on the noise levels in a noise condition of noise in a sound system, from the noise level to the noise conditions,.",Figure 1: Model structure for invariant training and ASR results.,0.7393,0.2528,0.496,596,542,1.1,55846ceac5b98a449e8481037d4f8893,images/2016/arxiv_0000165.png,https://arxiv.org/pdf/1612.01928v1.pdf
arxiv_0000166,Figure 166,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,Fig. 1. Example of an output of a visual active speaker detector.,A diagram of two men are standing in.,"A detailed layout showing two males in front of a curtain and a window with blinds behind them, one man wearing a hoodie.","A comprehensive technical explanation of a video game being played by two men in front of a window with curtains and windowsills behind them, with a man with a surprised expression on the camera taking a phone, and a man.",Fig. 1. Example of an output of a visual active speaker detector.,0.75,0.2435,0.4968,900,600,1.5,31d187efa45a8da8d9dfe360ac18cb29,images/2017/arxiv_0000166.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000167,Figure 167,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,Fig. 1. Example of an output of a visual active speaker detector.,A diagram of a man ' s face with the.,"A detailed layout showing the facial features of a man with a beard and mustache, with a positive and active face.","A comprehensive technical explanation of the facial recognition system for men with different facial types and abilitiess, including the facial hair color, the facial expression and the facial skin tone, the face shape, the ear, and the face.",Fig. 1. Example of an output of a visual active speaker detector.,0.81,0.1918,0.5009,407,164,2.482,18f9c53aae14589e633ad5588b9e0849,images/2017/arxiv_0000167.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000168,Figure 168,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,"Fig. 2. Approaches to visual active speaker detection considered in the study. In the first row are the perceptual inputs automatically extracted from the video
and audio streams. These inputs are passed to the task specific learning (second row), transfer learning (third row) and temporal learning (forth row) methods.",A diagram of the process for learning.,"A detailed layout showing the basic training cycle for instructional training and training materials, including audio track, task learning.","A comprehensive technical explanation of the virtual learning process for teachers and professionals of digital learning, including training, and training materials, and design, as well as described by the instructor, it is written by the following.","Fig. 2. Approaches to visual active speaker detection considered in the study. In the first row are the perceptual inputs automatically extracted from the video
and audio streams. These inputs are passed to the task specific learning (second row), transfer learning (third row) and temporal learning (forth row) methods.",0.81,0.3439,0.577,1062,618,1.718,7d31762e970464f9deacf71d70dec61e,images/2017/arxiv_0000168.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000169,Figure 169,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,"Fig. 4. Example of a difficult visual input from the first and second condition
in the dataset.",A diagram of a man with glasses and a.,"A detailed layout showing the man ' s beard and glasses from the reflection of his selfie camera that shows key features, attributes, and data.","A comprehensive technical explanation of the selfie of a man with glasses and a tie, selfie taken from a selfie camera, looking at the camera, with focus on his face, while taking a blurred image in the camera.","Fig. 4. Example of a difficult visual input from the first and second condition
in the dataset.",0.81,0.2404,0.5252,256,256,1.0,935e762d712dcaa711807be91840c0f0,images/2017/arxiv_0000169.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000170,Figure 170,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,NULL,A diagram of a woman with glasses.,A detailed layout showing the glasses and the hair of a lady with a phone in her hand and a cell phone in front of her face.,"A comprehensive technical explanation of the glasses used by the student to use the camera is not clear, but it ' s good to see if they are still working on the phones in the picture taken off the image without any.",NULL,0.75,0.3072,0.5286,256,256,1.0,654855c26a02f5903756fb77e189e1bf,images/2017/arxiv_0000170.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000171,Figure 171,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,NULL,A diagram of a man wearing a tie and.,"A detailed layout showing a man looking down at his cell phone, with the camera pointed up and his head tilted.","A comprehensive technical explanation of the procedure for hair transplants, including scalp and hair growth, and the treatment of hair loss and baldness, according for men with alopemys, or baldness or bald, bald.",NULL,0.7464,0.2678,0.5071,256,256,1.0,e9faa56cbc722d3341f779292cf2be22,images/2017/arxiv_0000171.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000172,Figure 172,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,"Fig. 4. Example of a difficult visual input from the first and second condition
in the dataset.",A diagram of three images showing the.,A detailed layout showing four images of a man wearing glasses and looking at the camera with a surprised look.,"A comprehensive technical explanation of the three images show how to use glasses for vision and vision recognition, including the eyes of a man with glasses on his head and a cell phone in his left side of his right hand and a neck.","Fig. 4. Example of a difficult visual input from the first and second condition
in the dataset.",0.7464,0.301,0.5237,584,304,1.921,84bf653ac83474f3e7ef68b93d399944,images/2017/arxiv_0000172.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000173,Figure 173,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,Fig. 3. Spatial configuration of the sensors and participants in the dataset.,A diagram of people around a table.,A detailed layout showing people sitting around a table with a large screen in the middle of them and hands grabbing around the edge.,"A comprehensive technical explanation of the table setting and how it works for you to use it in your office, home, or school room - youtube conference room, meeting room, or conference room - video conference room with people holding hands.",Fig. 3. Spatial configuration of the sensors and participants in the dataset.,0.81,0.287,0.5485,410,420,0.976,a4e98a0492a4708b1a3054398447d427,images/2017/arxiv_0000173.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000174,Figure 174,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,"Fig. 5. Accuracy versus participant and method. The participants are sorted by overall accuracy. The segment length for the LSTMs is 15 frames (500 ms).
The boxplots show the results over all 10 folds.",A diagram of the number and position.,"A detailed layout showing the number of cases per year in each county, and the growth of the number.","A comprehensive technical explanation of the global climate outlooks for the world from the end of 2012 to the end in 2018, with the trend of the year ending at least two years ending in red and the same period of the same.","Fig. 5. Accuracy versus participant and method. The participants are sorted by overall accuracy. The segment length for the LSTMs is 15 frames (500 ms).
The boxplots show the results over all 10 folds.",0.7071,0.2121,0.4596,1058,451,2.346,d842c9b8d334b9caa307454a9f1555b9,images/2017/arxiv_0000174.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000175,Figure 175,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,"Fig. 6. Comparison between audio-only and video-only method in noise (the solid lines are accuracies and the shaded areas are standard deviations). The
accuracy in the speaker dependent experiment (left) is averaged over 24 participants and 10 folds. The accuracy in the multi-speaker dependent experiment
(center) is averaged over 10 folds each containing data from 24 participants. The accuracy in the speaker independent experiment (right) is averaged over 24
folds each containing data from the participant that was left out during training. In all cases the performance of the audio-only method degrades with the
reduction in SNR, whereas the video-only method is not affected by acoustic noise.",A diagram of the average and depicting.,"A detailed layout showing the number of cases in each case, and the number on the other side that shows key features, attributes, and data points in.","A comprehensive technical explanation of the time series for a single - figure graphing function for each time period of a project, from the data center to the results table top of the data source, in the data table below,.","Fig. 6. Comparison between audio-only and video-only method in noise (the solid lines are accuracies and the shaded areas are standard deviations). The
accuracy in the speaker dependent experiment (left) is averaged over 24 participants and 10 folds. The accuracy in the multi-speaker dependent experiment
(center) is averaged over 10 folds each containing data from 24 participants. The accuracy in the speaker independent experiment (right) is averaged over 24
folds each containing data from the participant that was left out during training. In all cases the performance of the audio-only method degrades with the
reduction in SNR, whereas the video-only method is not affected by acoustic noise.",0.81,0.243,0.5265,1038,298,3.483,9990c8b2f2587a6168e3239204045732,images/2017/arxiv_0000175.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000176,Figure 176,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,"Jonas Beskow is a Professor in Speech Com-
munication with research interests in multimodal
speech technology, modeling and generating verbal
and non-verbal communicative behavior as well as
embodied conversational agents or social robots that
use speech, gesture and/or other modalities in order
to accomplish human-like interaction. He is also a
co-founder of Furhat Robotics, a startup developing
an innovative social robotics platform based on KTH
research.",A diagram of a man in glasses and a.,A detailed layout showing a man in a shirt and tie smiling for the camera while wearing glasses and a button down collared shirt.,"A comprehensive technical explanation of the role of a business development manager in the company, and how it is done to improve its operations and performance, with the client success of the teamwork on your own company, or a job.","Jonas Beskow is a Professor in Speech Com-
munication with research interests in multimodal
speech technology, modeling and generating verbal
and non-verbal communicative behavior as well as
embodied conversational agents or social robots that
use speech, gesture and/or other modalities in order
to accomplish human-like interaction. He is also a
co-founder of Furhat Robotics, a startup developing
an innovative social robotics platform based on KTH
research.",0.75,0.187,0.4685,120,160,0.75,bfc4991e664bc0f1f023ddeeea95fe78,images/2017/arxiv_0000176.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000177,Figure 177,scientific_figure,Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition,arXiv Research Authors,1711.08992v2,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Self-Supervised Vision-Based Detection of the Active Speaker as Support for Socially-Aware Language Acquisition. arXiv:1711.08992v2,"Kalin Stefanov received the MSc degree in Ar-
tificial Intelligence from University of Amsterdam
(Amsterdam, The Netherlands) and the PhD degree
in Computer Science from KTH Royal Institute
of Technology (Stockholm, Sweden). He is cur-
rently a post-doctoral fellow at the Institute for
Creative Technologies, University of Southern Cali-
fornia (Los Angeles, USA). His research interests in-
clude machine learning, computer vision and speech
technology.",A diagram of a man with a beard.,"A detailed layout showing a man with long hair and beard smiling for the camera, with a white background.","A comprehensive technical explanation of the basics of a professional photo shoot with a camera and a smile on his face, for example, a black and white photograph of a man with hair and beard and grey shirt and beard smiling.","Kalin Stefanov received the MSc degree in Ar-
tificial Intelligence from University of Amsterdam
(Amsterdam, The Netherlands) and the PhD degree
in Computer Science from KTH Royal Institute
of Technology (Stockholm, Sweden). He is cur-
rently a post-doctoral fellow at the Institute for
Creative Technologies, University of Southern Cali-
fornia (Los Angeles, USA). His research interests in-
clude machine learning, computer vision and speech
technology.",0.725,0.2404,0.4827,120,160,0.75,6674b97ee870b28ab3bfb7fbc621fe6a,images/2017/arxiv_0000177.png,https://arxiv.org/pdf/1711.08992v2.pdf
arxiv_0000178,Figure 178,scientific_figure,Sentiment Classification using Images and Label Embeddings,arXiv Research Authors,1712.00725v1,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Sentiment Classification using Images and Label Embeddings. arXiv:1712.00725v1,Figure 1: Data point example,A diagram of a cat yawning with its.,"A detailed layout showing the caption of a cat yawning on a chair with its mouth open that shows key features, attributes, and data points in detail.","A comprehensive technical explanation of the cat ' s behavior and behavior in a family pet, including a list of their features and their characteristics and characteristics, including its characteristics, as well known and language, and.",Figure 1: Data point example,0.81,0.2398,0.5249,1827,653,2.798,828334df9f570ba3028814ebadd26c0f,images/2017/arxiv_0000178.png,https://arxiv.org/pdf/1712.00725v1.pdf
arxiv_0000179,Figure 179,scientific_figure,Sentiment Classification using Images and Label Embeddings,arXiv Research Authors,1712.00725v1,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Sentiment Classification using Images and Label Embeddings. arXiv:1712.00725v1,Figure 1: Data point example,A diagram of a cat laying on a table.,"A detailed layout showing the text and images of an image of a cat with a red nose and tail that shows key features, attributes, and data points in.","A comprehensive technical explanation of the cat ' s diet, including its ingredients and how to use it for a healthy and nutrit eating diet, with a dog ' s owner, a pet - friendly environment, in a healthy.",Figure 1: Data point example,0.81,0.2465,0.5282,338,1346,0.251,7bf33c3f5cdb94d6ccb4af87f0d7ed4f,images/2017/arxiv_0000179.png,https://arxiv.org/pdf/1712.00725v1.pdf
arxiv_0000180,Figure 180,scientific_figure,Sentiment Classification using Images and Label Embeddings,arXiv Research Authors,1712.00725v1,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Sentiment Classification using Images and Label Embeddings. arXiv:1712.00725v1,Appendix 1: Model Comparison,A diagram of the average and depicting.,A detailed layout showing the number of different types of internet traffic in each country and their respective location.,"A comprehensive technical explanation of the best practices for agile software to use in the cloud - based business environment [ infographics ] ] % of the cloudworld, [ 1, 12 / 12, 2 / 3 % ] representing all details, specifications, and.",Appendix 1: Model Comparison,0.75,0.2005,0.4753,2255,866,2.604,35d06a2fc2b192a09feafe2a2f13ab0c,images/2017/arxiv_0000180.png,https://arxiv.org/pdf/1712.00725v1.pdf
arxiv_0000181,Figure 181,scientific_figure,Sentiment Classification using Images and Label Embeddings,arXiv Research Authors,1712.00725v1,cs,2017,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2017). Sentiment Classification using Images and Label Embeddings. arXiv:1712.00725v1,Appendix 2: Embedding Classifier Output in Sentiment label embedding space,A diagram of the different languages.,A detailed layout showing the different types of data generated by the user in the user ' s profile.,"A comprehensive technical explanation of the world map of the internet agen countries, by country and country of origin, in the world atlas or a4 / b / c / t / t s / t - t / w representing all details, specifications, and configurations of the.",Appendix 2: Embedding Classifier Output in Sentiment label embedding space,0.7071,0.2576,0.4823,1651,1590,1.038,0336abc225ecb2530e5f216f1b9ac068,images/2017/arxiv_0000181.png,https://arxiv.org/pdf/1712.00725v1.pdf
arxiv_0000182,Figure 182,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,"Figure 1. An example of a recurrent cell in our search space with 4 computational nodes. Left: The computational DAG that corresponds
to the recurrent cell. The red edges represent the flow of information in the graph. Middle: The recurrent cell. Right: The outputs of the
controller RNN that result in the cell in the middle and the DAG on the left. Note that nodes 3 and 4 are never sampled by the RNN, so
their results are averaged and are treated as the cell’s output.",A diagram of the process of making an.,"A detailed layout showing the process of producing a new cellphone for the us cellular network, with a full view of the cell phone.","A comprehensive technical explanation of the schemam for a single - cell cell phone system, showing the screeps and the screats in the scrab ' s position of the mains, and the same representing all details, specifications, and configurations of the.","Figure 1. An example of a recurrent cell in our search space with 4 computational nodes. Left: The computational DAG that corresponds
to the recurrent cell. The red edges represent the flow of information in the graph. Middle: The recurrent cell. Right: The outputs of the
controller RNN that result in the cell in the middle and the DAG on the left. Note that nodes 3 and 4 are never sampled by the RNN, so
their results are averaged and are treated as the cell’s output.",0.81,0.2466,0.5283,915,169,5.414,548e614b428ae45627efb461c7bf5854,images/2018/arxiv_0000182.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000183,Figure 183,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,"Figure 2. The graph represents the entire search space while the
red arrows define a model in the search space, which is decided
by a controller. Here, node 1 is the input to the model whereas
nodes 3 and 6 are the model’s outputs.",A diagram of the tree with three.,"A detailed layout showing the path of the two trees for each tree, including one of them missing that shows key features, attributes, and data.","A comprehensive technical explanation of the network tree for a tree - based approach to solve a problem in a system called a network tree, with a number of trees and numbers on each one side by side, and three, four.","Figure 2. The graph represents the entire search space while the
red arrows define a model in the search space, which is decided
by a controller. Here, node 1 is the input to the model whereas
nodes 3 and 6 are the model’s outputs.",0.81,0.2813,0.5456,332,187,1.775,04130eb4c4dbd989c3e52432a49069cf,images/2018/arxiv_0000183.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000184,Figure 184,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,"Figure 3. An example run of a recurrent cell in our search space
with 4 computational nodes, which represent 4 layers in a con-
volutional network. Top: The output of the controller RNN. Bot-
tom Left: The computational DAG corresponding to the network’s
architecture. Red arrows denote the active computational paths.
Bottom Right: The complete network. Dotted arrows denote skip
connections.",A diagram of the flow of data in an.,A detailed layout showing the process of using the model of the system to create a model for a real - time.,"A comprehensive technical explanation of the 3d - based model, including the model 3 and model 4, with the corresponding data and components labeled in figure 1 2, in figure 2, as well as described in figure, from the first.","Figure 3. An example run of a recurrent cell in our search space
with 4 computational nodes, which represent 4 layers in a con-
volutional network. Top: The output of the controller RNN. Bot-
tom Left: The computational DAG corresponding to the network’s
architecture. Red arrows denote the active computational paths.
Bottom Right: The complete network. Dotted arrows denote skip
connections.",0.7921,0.2781,0.5351,478,311,1.537,fbd88657155c06f4695b9d50ad9ecdab,images/2018/arxiv_0000184.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000185,Figure 185,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,"Figure 5. An example run of the controller for our search space
over convolutional cells. Top: the controller’s outputs. In our
search space for convolutional cells, node 1 and node 2 are the
cell’s inputs, so the controller only has to design node 3 and node
4. Bottom Left: The corresponding DAG, where red edges repre-
sent the activated connections. Bottom Right: the convolutional
cell according to the controller’s sample.",A diagram of the block - for - node.,"A detailed layout showing the block and layer of a system, including an array of blocks and ladders.",A comprehensive technical explanation of the block diagram for the layer 4 layer 3 layer 3 and layer 4 level 3 layer 4 layers 3 layer 5 layer 7 layer 3 layers 1 layer 3 block 4 layer 1 layer 5 layers 3 block.,"Figure 5. An example run of the controller for our search space
over convolutional cells. Top: the controller’s outputs. In our
search space for convolutional cells, node 1 and node 2 are the
cell’s inputs, so the controller only has to design node 3 and node
4. Bottom Left: The corresponding DAG, where red edges repre-
sent the activated connections. Bottom Right: the convolutional
cell according to the controller’s sample.",0.7671,0.2736,0.5203,478,570,0.839,dbde4a033f80dbe698d1c0777d063573,images/2018/arxiv_0000185.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000186,Figure 186,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,"Figure 4. Connecting 3 blocks, each with N convolution cells and
1 reduction cell, to make the final network.",A diagram of the sequence of an.,"A detailed layout showing the multi - step algorithm for adding numbers to a fraction system, using the following steps.","A comprehensive technical explanation of the multi - step algorithm for multiple - dimensional data retrieval and storage systems, which is used to determine the number of objects in each of the datas an array of the bits in the data system.","Figure 4. Connecting 3 blocks, each with N convolution cells and
1 reduction cell, to make the final network.",0.75,0.2891,0.5195,478,120,3.983,c1e80c16d75ba5628418564197a53b83,images/2018/arxiv_0000186.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000187,Figure 187,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,Figure 6. The RNN cell ENAS discovered for Penn Treebank.,A diagram of the three phases for the.,A detailed layout showing the functions of the three - dimensional system with a large number of elements and two different.,"A comprehensive technical explanation of thermos - based approach to thermo - based model for therms and thermocor - based systems, with three different functions in each of the same time, three - based.",Figure 6. The RNN cell ENAS discovered for Penn Treebank.,0.75,0.2512,0.5006,478,245,1.951,350007ee13f8eef809a9d543fad914b2,images/2018/arxiv_0000187.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000188,Figure 188,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,Figure 7. ENAS’s discovered network from the macro search space for image classification.,A diagram of the network that shows.,A detailed layout showing the different paths of a network with multiple connections in a single network and several connections.,"A comprehensive technical explanation of the algorithms for a networked networked system, including the following steps to create a network that displays each individual networked device in the same location and the networked connection of each.",Figure 7. ENAS’s discovered network from the macro search space for image classification.,0.81,0.2948,0.5524,740,279,2.652,7f7e18cd493bf8ed4abd1a6860eb62b6,images/2018/arxiv_0000188.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000189,Figure 189,scientific_figure,Efficient Neural Architecture Search via Parameter Sharing,arXiv Research Authors,1802.03268v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Efficient Neural Architecture Search via Parameter Sharing. arXiv:1802.03268v2,Figure 8. ENAS cells discovered in the micro search space.,A diagram of the cell line with.,A detailed layout showing the various processes of the process for cell differentiation and cellular structure in the cell.,"A comprehensive technical explanation of the first generation of the cell phone network, with multiple cellular connections and cellular structure diagram below it is labeled in the following diagram for the image below, rljtd - t1,.",Figure 8. ENAS cells discovered in the micro search space.,0.75,0.2688,0.5094,469,763,0.615,10f36509c0bb6061e415d71b12e332c3,images/2018/arxiv_0000189.png,https://arxiv.org/pdf/1802.03268v2.pdf
arxiv_0000190,Figure 190,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 1: Illustrative representation of the problem with the
original product data – most products have part of the at-
tributes missing and almost no products have all the at-
tributes available.",A diagram of the words in the poem are.,"A detailed layout showing the words in the top row of the text book, which are different colors that shows key features, attributes, and data points.","A comprehensive technical explanation of the basic pattern for clothing and shoes, for women and men, as well as their characteristics and uses, in a texturally, and an example, a dress, a, a - - representing all details, specifications, and.","Figure 1: Illustrative representation of the problem with the
original product data – most products have part of the at-
tributes missing and almost no products have all the at-
tributes available.",0.81,0.3176,0.5638,565,319,1.771,e25d98aebc52222252f041bc4f7163de,images/2018/arxiv_0000190.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000191,Figure 191,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 2: Example instances with their associated (a) product images and (b) product details, as they appear on product pages.
Using these inputs together with a product’s type, brand and division, the network predicts a set of product attributes (c).",A diagram of the different types of.,"A detailed layout showing the different styles of dresses in various colors and sizes, with text on them.","A comprehensive technical explanation of the fashions of the models on the spring / summer 2015 collection, including a short - sleeved dress and long - sleeve top with a fitted jacket, skinny - length, low - fitting,.","Figure 2: Example instances with their associated (a) product images and (b) product details, as they appear on product pages.
Using these inputs together with a product’s type, brand and division, the network predicts a set of product attributes (c).",0.725,0.3023,0.5136,1224,769,1.592,43e84522a80fa26ee76b411c8a260e1e,images/2018/arxiv_0000191.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000192,Figure 192,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 3: Multi-modal multi-task architecture. In grey the
inputs to the network, in blue the hidden layers and in red
the output layers (one per task, i.e., attribute). In parenthesis,
the output size of each layer, which reflects the input encod-
ing for input layers, the embedding dimensionality for em-
bedding layers, and the number of filters and neurons for
convolutional and dense layers, respectively.",A diagram of the blockchaik model with.,"A detailed layout showing the different types of the software used for the application of the system, including the cpu.","A comprehensive technical explanation of the hpl - s / t model for the internet network, including the network layer and the network level structure, as described by the host layer, and host layer and host, and route layer.","Figure 3: Multi-modal multi-task architecture. In grey the
inputs to the network, in blue the hidden layers and in red
the output layers (one per task, i.e., attribute). In parenthesis,
the output size of each layer, which reflects the input encod-
ing for input layers, the embedding dimensionality for em-
bedding layers, and the number of filters and neurons for
convolutional and dense layers, respectively.",0.75,0.2649,0.5074,517,414,1.249,4ca911b2e5a0510a9b226c0f1a3e8087,images/2018/arxiv_0000192.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000193,Figure 193,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 4: Evolution of the weighted F1 scores, measured at in-
tervals of 1000 steps per attribute (minibatch size of 64). The
model achieves an average weighted f1 of 0.855. The score is
still increasing in train but plateaus out in test. per attribute. The confusion matrices for two of the attributes are
shown in Figure 5. We can see that for neckline the mass is heavily
concentrated on the diagonal. On the off-diagonal, the most signifi-
cant mismatch is between ‘button collar’ and ‘point collar’, both
types of shirts neckline. Samples from the class which has the least
support (‘point collar’), are often misclassified as ‘button collar’,
which is five times more frequent. The use/occasion attribute repre-
sents a higher-level categorization that is harder to capture from the
data; still, the performance is way above chance level with a 0.665
accuracy (vs 0.238 baseline, 13-class task), with ‘wedding’ being the
easiest class to spot. The average of the weighted F1 scores is 0.855
when using all inputs (image, text, metadata). The most important
input is text: without it, the F1 score drops to 0.773. Images are
the second most important: without them, the F1 score drops to
0.842. Finally, metadata is the least important: without it, the score
drops to 0.849. The relative importance of image and text is in line
with what has been observed in a related problem and modelling
approach [29], where a text CNN alone can classify e-commerce
products better than an image CNN alone, and multi-modal fusion
only brings a marginal improvement. The impact of removing each
input can be seen in more detail in Figure 6. 5",A diagram of a graph with squares on.,"A detailed layout showing the steps of a blue and white mosaic pattern, with a few squares to be seen.","A comprehensive technical explanation of the graph of growth in blue squares on white background, with text below it that reads, ' the graph is going to be a step up ' s the number of growth ' s up ' '.","Figure 4: Evolution of the weighted F1 scores, measured at in-
tervals of 1000 steps per attribute (minibatch size of 64). The
model achieves an average weighted f1 of 0.855. The score is
still increasing in train but plateaus out in test. per attribute. The confusion matrices for two of the attributes are
shown in Figure 5. We can see that for neckline the mass is heavily
concentrated on the diagonal. On the off-diagonal, the most signifi-
cant mismatch is between ‘button collar’ and ‘point collar’, both
types of shirts neckline. Samples from the class which has the least
support (‘point collar’), are often misclassified as ‘button collar’,
which is five times more frequent. The use/occasion attribute repre-
sents a higher-level categorization that is harder to capture from the
data; still, the performance is way above chance level with a 0.665
accuracy (vs 0.238 baseline, 13-class task), with ‘wedding’ being the
easiest class to spot. The average of the weighted F1 scores is 0.855
when using all inputs (image, text, metadata). The most important
input is text: without it, the F1 score drops to 0.773. Images are
the second most important: without them, the F1 score drops to
0.842. Finally, metadata is the least important: without it, the score
drops to 0.849. The relative importance of image and text is in line
with what has been observed in a related problem and modelling
approach [29], where a text CNN alone can classify e-commerce
products better than an image CNN alone, and multi-modal fusion
only brings a marginal improvement. The impact of removing each
input can be seen in more detail in Figure 6. 5",0.7143,0.2778,0.496,549,549,1.0,b2220692a1653b76ec29deeebfbf0641,images/2018/arxiv_0000193.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000194,Figure 194,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 5: Examples of test-set normalised confusion matri-
ces for target attributes; true labels (by row) vs predicted la-
bels (by column). Attributes might be related to a product
shape or other visual characteristics (a), but also to higher-
level properties that are not directly expressed in the prod-
uct data (b). Regarding the former, although the model does
well overall, one can still observe biases towards the most
popular classes such as ‘crew’ neck and ‘standard’ neck. Re-
garding the latter, the easiest end use to predict is ‘wedding’
whereas the most challenging is ‘going out bar’, which is of-
ten confused with ‘evening occasion’.",A diagram of a blue and white pattern.,A detailed layout showing the blue squares on the right side of the image are arranged in a straight line.,"A comprehensive technical explanation of the future of the internet market, including the internet and its capabilities for business and social use, from the internet to the internet age of the web age of today ' s technology era era.","Figure 5: Examples of test-set normalised confusion matri-
ces for target attributes; true labels (by row) vs predicted la-
bels (by column). Attributes might be related to a product
shape or other visual characteristics (a), but also to higher-
level properties that are not directly expressed in the prod-
uct data (b). Regarding the former, although the model does
well overall, one can still observe biases towards the most
popular classes such as ‘crew’ neck and ‘standard’ neck. Re-
garding the latter, the easiest end use to predict is ‘wedding’
whereas the most challenging is ‘going out bar’, which is of-
ten confused with ‘evening occasion’.",0.7286,0.293,0.5108,441,441,1.0,2079ad2cf7bb37265200b09eb30e4f1c,images/2018/arxiv_0000194.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000195,Figure 195,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 4: Evolution of the weighted F1 scores, measured at in-
tervals of 1000 steps per attribute (minibatch size of 64). The
model achieves an average weighted f1 of 0.855. The score is
still increasing in train but plateaus out in test.",A diagram of the average age and age.,"A detailed layout showing the average and glyclia of all athletes in different groups, including runners.","A comprehensive technical explanation of average and glss of all athletes, 2004 - 2010, with data by age and gender, 2006 - 2009, and 2010 - 2012 - 2011 - 2013, and 2009 - 2015, and the same representing all details, specifications, and.","Figure 4: Evolution of the weighted F1 scores, measured at in-
tervals of 1000 steps per attribute (minibatch size of 64). The
model achieves an average weighted f1 of 0.855. The score is
still increasing in train but plateaus out in test.",0.725,0.2548,0.4899,491,613,0.801,9ba08fe7d6c5f3c73158b5e7c89b1741,images/2018/arxiv_0000195.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000196,Figure 196,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 5: Examples of test-set normalised confusion matri-
ces for target attributes; true labels (by row) vs predicted la-
bels (by column). Attributes might be related to a product
shape or other visual characteristics (a), but also to higher-
level properties that are not directly expressed in the prod-
uct data (b). Regarding the former, although the model does
well overall, one can still observe biases towards the most
popular classes such as ‘crew’ neck and ‘standard’ neck. Re-
garding the latter, the easiest end use to predict is ‘wedding’
whereas the most challenging is ‘going out bar’, which is of-
ten confused with ‘evening occasion’.",A diagram of a series of graphs.,A detailed layout showing the number of different types of aircraft that are flying in the sky and below.,"A comprehensive technical explanation of the evolution of the internet web in the past and present infographics of the past, from wikipediart to the present time of the future source source info / source / source - info / info.","Figure 5: Examples of test-set normalised confusion matri-
ces for target attributes; true labels (by row) vs predicted la-
bels (by column). Attributes might be related to a product
shape or other visual characteristics (a), but also to higher-
level properties that are not directly expressed in the prod-
uct data (b). Regarding the former, although the model does
well overall, one can still observe biases towards the most
popular classes such as ‘crew’ neck and ‘standard’ neck. Re-
garding the latter, the easiest end use to predict is ‘wedding’
whereas the most challenging is ‘going out bar’, which is of-
ten confused with ‘evening occasion’.",0.725,0.2367,0.4808,526,980,0.537,e4e51a0f0cf4917d830dba4dcdce581d,images/2018/arxiv_0000196.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000197,Figure 197,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 6: Ablation study. Density distribution (y-axis) of test
F1 scores (x-axis) across all predicted attributes when re-
moving one of the inputs in turns (see legend). The average
F1 scores are: 0.855 for the full model, 0.849 without prod-
uct metadata (product type, brand, and division), then 0.842
without image features, and 0.773 without text input.",A diagram of the distribution of.,A detailed layout showing the distribution of the number of samples of data for each data source in the data.,"A comprehensive technical explanation of the data flow in the data visual system, including the data processing process and the data data processing software and its capabilities on the webpages and the web interfaces on the desktops screen -.","Figure 6: Ablation study. Density distribution (y-axis) of test
F1 scores (x-axis) across all predicted attributes when re-
moving one of the inputs in turns (see legend). The average
F1 scores are: 0.855 for the full model, 0.849 without prod-
uct metadata (product type, brand, and division), then 0.842
without image features, and 0.773 without text input.",0.7993,0.2726,0.536,442,291,1.519,05cae790900453325de97a56492ea52c,images/2018/arxiv_0000197.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000198,Figure 198,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 7: Hybrid recommender architecture. Content-based
(CB) and collaborative (CF) inputs to the network are de-
picted in grey, hidden layers in blue, functional layers in pur-
ple, and final layer in red. In parenthesis, the output size of
each layer: #a denotes the dimensionality of the encoded at-
tributes, z the number of negative samples per batch in the
ranking loss, and k the dimension of the product factors.",A diagram of the block structure for a.,"A detailed layout showing the basic components of a computer system with multiple functions and functions, including a cpu.","A comprehensive technical explanation of the architecture of a multi - platform system, including a central processor, a virtual machine and a virtual device, including an appliancepter, a desktop, a webinator,.","Figure 7: Hybrid recommender architecture. Content-based
(CB) and collaborative (CF) inputs to the network are de-
picted in grey, hidden layers in blue, functional layers in pur-
ple, and final layer in red. In parenthesis, the output size of
each layer: #a denotes the dimensionality of the encoded at-
tributes, z the number of negative samples per batch in the
ranking loss, and k the dimension of the product factors.",0.75,0.2525,0.5012,460,438,1.05,b79710594ac297ac2c9a9c743fa5415f,images/2018/arxiv_0000198.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000199,Figure 199,scientific_figure,Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products,arXiv Research Authors,1803.07679v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Product Characterisation towards Personalisation: Learning Attributes from Unstructured Data to Recommend Fashion Products. arXiv:1803.07679v1,"Figure 8: Examples of similar products defined by the recommendation models. Top row: product with the largest number of
customer interactions in the training set. Middle row: product with a median number of customer interactions. Bottom row:
least popular product in the training period. We can see that collaborative performs better for products with more interactions,
while content and hybrid are not affected by this factor.",A diagram of various styles of men ' s.,A detailed layout showing the various types of bags that are available for purchase on the website of the store.,"A comprehensive technical explanation of the different types of bags and their uses for travel, featuring what they are meant to be used in the world, and where they are made, and what they look good, and how they.","Figure 8: Examples of similar products defined by the recommendation models. Top row: product with the largest number of
customer interactions in the training set. Middle row: product with a median number of customer interactions. Bottom row:
least popular product in the training period. We can see that collaborative performs better for products with more interactions,
while content and hybrid are not affected by this factor.",0.75,0.361,0.5555,1224,617,1.984,8150a0712a9fb86ee435f2ad259dba58,images/2018/arxiv_0000199.png,https://arxiv.org/pdf/1803.07679v1.pdf
arxiv_0000200,Figure 200,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 1: Context Aware Captioner. At each step t, the
textual information w∗
t−1, and the mixture of image features
and visual sentinel ˆct−1 from previous step t−1 are fed to
the LSTM to make it aware of past attentional contexts.",A diagram of the algorithm for the.,"A detailed layout showing the path of the lml and the lmt functions in the system, from which that shows key features, attributes, and data points.","A comprehensive technical explanation of the algorithm for the linear system, from the textbook of systems and applications, by john j s davis, m d, and j l e w, r d, d, l, m representing all details, specifications, and configurations of the.","Figure 1: Context Aware Captioner. At each step t, the
textual information w∗
t−1, and the mixture of image features
and visual sentinel ˆct−1 from previous step t−1 are fed to
the LSTM to make it aware of past attentional contexts.",0.81,0.3262,0.5681,230,338,0.68,7e4f00d21c8f1568d4cc9d6c0e4744a9,images/2018/arxiv_0000200.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000201,Figure 201,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 2: Discriminator architectures. (a) Joint-Embedding Discriminator from [4]. (b) Our proposed Dη. By jointly
embedding the image and caption with a co-attention model, we give the discriminator the ability to modulate the image
features depending on the caption and vice versa.",A diagram of the basic system for the.,"A detailed layout showing the various parts of a computer system, including an atm / atm and atm that shows key features, attributes, and data.","A comprehensive technical explanation of the system for the control of electronic devices, including the control system, and the control systems, including a control unit, control system and control system for each of the control engine, control.","Figure 2: Discriminator architectures. (a) Joint-Embedding Discriminator from [4]. (b) Our proposed Dη. By jointly
embedding the image and caption with a co-attention model, we give the discriminator the ability to modulate the image
features depending on the caption and vice versa.",0.81,0.2263,0.5181,889,340,2.615,b0adb82a37e29b1ba837ce8404e15ec3,images/2018/arxiv_0000201.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000202,Figure 202,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 3: SCST Training of GAN-captioning.,A diagram of a complex system with a.,"A detailed layout showing the complex of a system with different functions and functions, including the following function.","A comprehensive technical explanation of the fourier theory of the system for a complex function of a function of the function of an object in a series of functions, c - p, p, d = d, d, t representing all details, specifications, and configurations.",Figure 3: SCST Training of GAN-captioning.,0.75,0.2415,0.4958,446,166,2.687,b0af6cc82ba19073dc66c357346026ec,images/2018/arxiv_0000202.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000203,Figure 203,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 4: Evolution of semantic scores over training epochs for COCO Test and OOC datasets. Our Co-att models achieve
consistently higher scores than CE, RL and Joint-Emb models [4].",A diagram of the two plots show.,A detailed layout showing the relationship of the different levels of the data for the data and the data.,"A comprehensive technical explanation of the initial data of the world ' s largest seawater and ocean water cycle, from the u s - china sea level and the pacific ocean, to the north pole, and the atlantic, and atlantic.","Figure 4: Evolution of semantic scores over training epochs for COCO Test and OOC datasets. Our Co-att models achieve
consistently higher scores than CE, RL and Joint-Emb models [4].",0.725,0.2547,0.4899,704,243,2.897,7035e6f7fcbe9a6db96a15720245f4d8,images/2018/arxiv_0000203.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000204,Figure 204,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 5: Evolution of vocabulary coverage over training epochs for COCO and OOC datasets. As training progresses, we
see a correlation between vocabulary coverage and semantic scores for all models. Models without CIDEr-regularized SCST
GAN rewards achieve best vocabulary coverage.",A diagram of three different lines.,"A detailed layout showing the different types of data in each chart each of them is a single line that shows key features, attributes, and data.","A comprehensive technical explanation of the differences between the three trading ranges of the stock market and the stock price of the same country of the country of china, including the world - china and the united states of the world.","Figure 5: Evolution of vocabulary coverage over training epochs for COCO and OOC datasets. As training progresses, we
see a correlation between vocabulary coverage and semantic scores for all models. Models without CIDEr-regularized SCST
GAN rewards achieve best vocabulary coverage.",0.87,0.2552,0.5626,718,243,2.955,157c35d0c54b2048edc6d8fa007c1d9a,images/2018/arxiv_0000204.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000205,Figure 205,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 6: Human evaluations of EnsCE, EnsRL and several GAN ensembles on COCO and OOC sets. (a) A distribution of
preferences for the best caption determined by the majority of 5 human evaluators; here GAN label indicates the Ens2 model.
(b) Turing test on detecting the human-written versus GAN-generated captions on COCO. We assign ""yes/no’ with at least 4
out of 5, disagree otherwise. (c) Mean opinion score vs. Semantic score on COCO test images.",A diagram of the different types of.,"A detailed layout showing the different types of the data that can be used to create a data visual that shows key features, attributes, and data.","A comprehensive technical explanation of the different data processing techniques used in the machine to produce data for the system, including the number of datas generated by the computer and the user ' s choice of the system ' s data,.","Figure 6: Human evaluations of EnsCE, EnsRL and several GAN ensembles on COCO and OOC sets. (a) A distribution of
preferences for the best caption determined by the majority of 5 human evaluators; here GAN label indicates the Ens2 model.
(b) Turing test on detecting the human-written versus GAN-generated captions on COCO. We assign ""yes/no’ with at least 4
out of 5, disagree otherwise. (c) Mean opinion score vs. Semantic score on COCO test images.",0.81,0.231,0.5205,900,264,3.409,001c7dc4c19970be8aa810372013e125,images/2018/arxiv_0000205.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000206,Figure 206,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,A diagram of a street with parked.,A detailed layout showing a narrow alley with parked motorcycles in front of the building and a person walking.,"A comprehensive technical explanation of a motorcycle parked in an alleyway in india, with people walking by the alleyway and motorcycles parked on the side of the street in front of the alleys and behind the buildings, a man.",Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,0.7464,0.3358,0.5411,300,300,1.0,aa8a33b1af9cc767929ebd671eb5d224,images/2018/arxiv_0000206.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000207,Figure 207,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of two giraffes standing in.,"A detailed layout showing two giraffes standing in a field of grass in front of a tree that shows key features, attributes, and data points in.","A comprehensive technical explanation of giraffes in the wild for beginners to learn how to eat them and why to use their food to feed them for them to survive and how to survive them to help them, including.",NULL,0.81,0.2977,0.5539,300,300,1.0,12aa8e41336cc0b9d05a4f7526a1ac4e,images/2018/arxiv_0000207.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000208,Figure 208,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 8: L2 norm of the gradient with respect to the log-
its during training of Gθ with different training strategies.
The plots show a minibatch-mean during the training; the
variance of each curve gives a good idea of the gradient
stability between minibatches. We can see that SCST with
pure discriminator reward has the lowest gradient norm.",A diagram of a bus traveling down the.,"A detailed layout showing the street and the traffic on one side of the road, and the bus on the other side of it.","A comprehensive technical explanation of the bus on the road, and how to use it for transportation purposes, according with the following rules and guidelines, as well as well, as described by a driver or by person who is the driver,.","Figure 8: L2 norm of the gradient with respect to the log-
its during training of Gθ with different training strategies.
The plots show a minibatch-mean during the training; the
variance of each curve gives a good idea of the gradient
stability between minibatches. We can see that SCST with
pure discriminator reward has the lowest gradient norm.",0.75,0.2857,0.5179,300,300,1.0,0932bb0935d2df5b3b41528b8a97c515,images/2018/arxiv_0000208.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000209,Figure 209,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,A diagram of a stop sign with graffiti.,A detailed layout showing the stop sign with graffiti on it and a spray can next to it with a sticker on it.,"A comprehensive technical explanation of stop the red light sign, with spray paint on it ' s sides and a sticker on the side of it that reads you cant can ' t stop the rest in the reddx representing all details, specifications, and configurations.",Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,0.7357,0.3394,0.5375,300,300,1.0,80fe28a8c79fc706004cc63fe25e47b3,images/2018/arxiv_0000209.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000210,Figure 210,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a bird sitting on the.,"A detailed layout showing a bird and a laptop on the floor with confetti scattered around it that shows key features, attributes, and data points in.","A comprehensive technical explanation of parrots and their problems in the home office environment - page 2 - image jp _ _ _ 0078 _ _ png jp2 _ c4 _ 1 jp _ wm jpg representing all details, specifications, and configurations of the system components.",NULL,0.81,0.3488,0.5794,300,300,1.0,8cf1ad7ce8fc0f9d916272ec1b0b3556,images/2018/arxiv_0000210.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000211,Figure 211,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a car that is sitting on.,A detailed layout showing the front end of the car on the street and the driver standing next to it.,A comprehensive technical explanation of a car accident involving a vehicle on the road and a person standing near it with a skateboard on the ground in front of it and a car that is sitting on a bench on the side of the road.,NULL,0.7071,0.2396,0.4733,300,300,1.0,8e84bed0f54c27eae3f44eb3d1749f5e,images/2018/arxiv_0000211.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000212,Figure 212,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,A diagram of a person crossing a.,A detailed layout showing a man balancing on a stop sign and people riding bikes in the background with a building in the distance.,"A comprehensive technical explanation of a stop sign with a man on a bicycle in front of a building and several cars parked on the street behind it, with people on bikes in front and on bicycles in the foreground, in the background.",Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,0.75,0.2939,0.522,300,300,1.0,d6cc99f95b8bb441e73012a3c6b104e8,images/2018/arxiv_0000212.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000213,Figure 213,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a street light with a.,A detailed layout showing the street light and the large potted planter on the corner of the street.,A comprehensive technical explanation of the street lamp in a city area with tall buildings and a man walking past it on a sidewalk in front of a tall building with a tall orange lightpost with a blue sign that says transcount.,NULL,0.7071,0.2905,0.4988,300,300,1.0,a58188aa721451330f6abc7c3d2b6012,images/2018/arxiv_0000213.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000214,Figure 214,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,A diagram of different signs and.,A detailed layout showing several pictures of animals and cars on the street with text in english and spanish.,"A comprehensive technical explanation of the different types of traffic signs in the world, including stop sign, deer, zebra, and motorcycle, and various other vehicles with images of a car and people on the same street scene in the same.",Figure 7: Examples of captions for our proposed model on COCO and OOC sets.,0.7429,0.3464,0.5446,730,634,1.151,8851182e207ff0f0101455a246c36342,images/2018/arxiv_0000214.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000215,Figure 215,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 8: L2 norm of the gradient with respect to the log-
its during training of Gθ with different training strategies.
The plots show a minibatch-mean during the training; the
variance of each curve gives a good idea of the gradient
stability between minibatches. We can see that SCST with
pure discriminator reward has the lowest gradient norm.",A diagram of the different types of.,"A detailed layout showing the number of different types of sound waves in the same band of sound, and the number in each.","A comprehensive technical explanation of the current and future spectrums in the system of the same time as well as the actual time of the signal source source of the source, according the source and the source of this image,.","Figure 8: L2 norm of the gradient with respect to the log-
its during training of Gθ with different training strategies.
The plots show a minibatch-mean during the training; the
variance of each curve gives a good idea of the gradient
stability between minibatches. We can see that SCST with
pure discriminator reward has the lowest gradient norm.",0.75,0.2108,0.4804,455,285,1.596,0a0cdb6bc3bea1aa81da407ee0f4c1e9,images/2018/arxiv_0000215.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000216,Figure 216,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 9: Semantic vs. Discriminator scores across 40 training epochs for ground truth captions using the joint embedding
discriminator [4].",A diagram of several plots with.,A detailed layout showing the various distribution of each type of data in a ploter sheet with different colors.,"A comprehensive technical explanation of the scatter plot for the different type of data visual data visual, data visual visual, visual data, data analysis, data data visual and visual, graph visual, scatter, visual, information visual,.","Figure 9: Semantic vs. Discriminator scores across 40 training epochs for ground truth captions using the joint embedding
discriminator [4].",0.87,0.2912,0.5806,610,739,0.825,e5e3c35fe378cb7bf4d9f8c4aa3b2408,images/2018/arxiv_0000216.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000217,Figure 217,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 9: Semantic vs. Discriminator scores across 40 training epochs for ground truth captions using the joint embedding
discriminator [4].",A diagram of a number of different.,"A detailed layout showing the different color schemes for a game of doting games, including a number of dots.","A comprehensive technical explanation of the scatter plot of a large number of different types of datas and their corresponding elements, including data visual, data visual and visual, image, and data, and visual and time,.","Figure 9: Semantic vs. Discriminator scores across 40 training epochs for ground truth captions using the joint embedding
discriminator [4].",0.7393,0.2242,0.4818,1224,1111,1.102,68355b2fe15fb5260c1a8dc402bc8bb1,images/2018/arxiv_0000217.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000218,Figure 218,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 10: The interface of ""Fine Grained Evaluation"".",A diagram of a woman playing tennis.,"A detailed layout showing the contents of a lesson for children to learn tennis, including a lesson on court.","A comprehensive technical explanation of the four part sports court exercise - page 9 of 8 - with answers included - click to enlarge image below to env, click to see full size version in the above gallery,.","Figure 10: The interface of ""Fine Grained Evaluation"".",0.7393,0.2511,0.4952,1420,1338,1.061,aa7bce3a7f9409fc64c4aee33feb4b9b,images/2018/arxiv_0000218.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000219,Figure 219,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,"Figure 10: The interface of ""Fine Grained Evaluation"".",A diagram of a girl playing tennis.,A detailed layout showing the correct spelling and numbers for this picture of a girl playing tennis with her dad.,A comprehensive technical explanation of the tennis court and the woman playing tennis on a court with a child in the background and the number 1 on the court below the score sheet of the score is the ball and the player with the score.,"Figure 10: The interface of ""Fine Grained Evaluation"".",0.75,0.3176,0.5338,518,908,0.57,e4862edadaecd079a9f0f5b8259c7b18,images/2018/arxiv_0000219.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000220,Figure 220,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 11: Discrimator scores across different training Gumbel Methods.,A diagram of the same temperature and.,A detailed layout showing the different types of data generated by various data sources in the same data source.,"A comprehensive technical explanation of the data visualization for the data processing system, using the data management software and the data modeling tool, with the data driven - based on - based image of the same datas and the same.",Figure 11: Discrimator scores across different training Gumbel Methods.,0.75,0.2355,0.4928,959,643,1.491,d599b7858ff1fa7dcf6b91cc57bbc2bf,images/2018/arxiv_0000220.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000221,Figure 221,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a line graph showing the.,"A detailed layout showing the data of the two different graphs for each time period of the day, as well as the following one.","A comprehensive technical explanation of the average timeframe for the data of the two different systems, and how each one is different from the previous model, based on the previous timeframes? - based on timeframe??.",NULL,0.81,0.2883,0.5492,984,642,1.533,c0833c84a19765779bbeb2ace2d3ba83,images/2018/arxiv_0000221.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000222,Figure 222,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of the data displayed on a.,"A detailed layout showing the data of the two different graphs, with different colors and sizes of the data.","A comprehensive technical explanation of the results of the two - day data analysis on the same line graph, including the data and data visual data visualization of the data visual for each time data source for the data in the data.",NULL,0.7957,0.273,0.5343,990,642,1.542,9c56c5b20c52b19729f404bd24d5e9e9,images/2018/arxiv_0000222.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000223,Figure 223,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 11: Discrimator scores across different training Gumbel Methods.,A diagram of a line graph showing the.,A detailed layout showing the distribution of the current temperature and the current heat patterns of the two regions.,"A comprehensive technical explanation of the temperature curve, and the temperature change data for the future year, and how to use the data to measure it?? - page 2 /??? /? / /? - - representing all details, specifications, and configurations of.",Figure 11: Discrimator scores across different training Gumbel Methods.,0.81,0.2168,0.5134,663,279,2.376,34356091f57fac74480a9e2c909f89f4,images/2018/arxiv_0000223.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000224,Figure 224,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 12: Cherry-picked examples on the COCO validation set.,A diagram of a small dock with several.,A detailed layout showing several boats docked in the water at the marina area of a resort or resort.,"A comprehensive technical explanation of the lake and its surroundings, including the boats, grass, and trees, and the water ' s surface in the foregrounds, is clear water and the lake, and blue sky with clouds.",Figure 12: Cherry-picked examples on the COCO validation set.,0.7107,0.289,0.4999,300,300,1.0,d577c78cff53df16f9d42b55fb99fb67,images/2018/arxiv_0000224.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000225,Figure 225,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of the man is hitting a.,"A detailed layout showing a man hitting a baseball with a bat in the dark night sky, and spray spraying from his mouth.","A comprehensive technical explanation of the best baseball bat for hitting a ball with a bat in his hand, and how to use it at night to hit it? - out a pitch the perfect for a game in the game representing all details, specifications, and.",NULL,0.75,0.3332,0.5416,300,300,1.0,e9883c8692b8fc076201ca17285eff26,images/2018/arxiv_0000225.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000226,Figure 226,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a store window with.,A detailed layout showing a variety of umbrellas and other items in a store window display area with a horse.,"A comprehensive technical explanation of shop displays for umbrellas in a store window display window displays, display design, retail displays, retail ideas, display windows, window displays and display design ideas, retail design, display.",NULL,0.7393,0.3097,0.5245,300,300,1.0,bc4b031d2f18d4105c763fd776f9c703,images/2018/arxiv_0000226.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000227,Figure 227,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 12: Cherry-picked examples on the COCO validation set.,A diagram of the four images show how.,A detailed layout showing four different images in black and white of people playing with toys and water guns.,"A comprehensive technical explanation of the various types of photography used in this article, including photos and texting, and video editing, for each subject or subject ' s purpose ' s own purpose ' twep ' s thoughts.",Figure 12: Cherry-picked examples on the COCO validation set.,0.7429,0.3135,0.5282,595,1244,0.478,f80b5e062d0c5a859d175e463b3ac147,images/2018/arxiv_0000227.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000228,Figure 228,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 13: Cherry-picked examples on the Out of Context (OOC) set.,A diagram of the damage from a car in.,A detailed layout showing the damaged glass of a store front and a car in the foreground of the shop.,"A comprehensive technical explanation of the damage from a car that crashed into a chinese restaurant in the city of london, england, on january 19, 2012, 2012 - - - by - tom hengensen / - tom hu representing all details, specifications, and.",Figure 13: Cherry-picked examples on the Out of Context (OOC) set.,0.7107,0.3381,0.5244,300,300,1.0,53009a7e89d0e769868a6f43f963bcb2,images/2018/arxiv_0000228.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000229,Figure 229,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a bed in the middle of a.,A detailed layout showing a bed in the middle of a forest with green grass and tall trees in the background.,"A comprehensive technical explanation of a bed in the woods and how to use it for sleep and rest on a low - lying bed, or a large bed made bed with a mattress, frame, in the middle, and a frame representing all details, specifications, and.",NULL,0.7357,0.3216,0.5287,300,300,1.0,5acc652c394c719ccd2e28433fa3a0b4,images/2018/arxiv_0000229.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000230,Figure 230,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a couch in the middle of.,"A detailed layout showing a couch and trash can in the street near a garage and a trash can that shows key features, attributes, and data points in.",A comprehensive technical explanation of a couch sitting in the grass on the side of the street with garbage cans nearby and a garage door behind it and a trash can next to the curb with a white garage door and a green lawn.,NULL,0.81,0.3235,0.5668,300,300,1.0,32c48a9a11497fd823bb628a74458f56,images/2018/arxiv_0000230.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000231,Figure 231,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of the airport is very busy.,A detailed layout showing a plane landing over a traffic jam on the road with cars waiting to be stopped.,"A comprehensive technical explanation of the airport and its surrounding environment, including the runway and the landing strip, is important to any passenger who travels in the area, according, especially on the runway, and at the airport or.",NULL,0.725,0.3464,0.5357,300,300,1.0,1166ce61e8ac5e0ce4f172db6818ce50,images/2018/arxiv_0000231.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000232,Figure 232,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 13: Cherry-picked examples on the Out of Context (OOC) set.,A diagram of different types of.,"A detailed layout showing the various types of the building and its surroundings, including a parking lot, a car garage, and a store.","A comprehensive technical explanation of the different types of pipes used for transportation and storage in the city of vancouver, canada, including the gas pump, the car washer, the new one that was built to the fire station,.",Figure 13: Cherry-picked examples on the Out of Context (OOC) set.,0.75,0.2804,0.5152,635,1252,0.507,3a3632c1cdf1472c24cc5274e19950a6,images/2018/arxiv_0000232.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000233,Figure 233,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 14: Lime-picked examples on the COCO test set.,A diagram of a baseball pitcher.,A detailed layout showing pitcher throwing a baseball from the mound in a stadium setting with people watching from bleachers.,"A comprehensive technical explanation of the texas rangers ' starting pitcher, including the pitching motion and the pitch that will be thrown to him in the first inning on may 22, 2012, at the mound at the aldsup photo.",Figure 14: Lime-picked examples on the COCO test set.,0.75,0.2394,0.4947,300,300,1.0,5776eb4d33d815eb819a9e622b22a022,images/2018/arxiv_0000233.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000234,Figure 234,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a man and woman posing.,"A detailed layout showing the portrait of a man and woman in an old photo, posing for a picture that shows key features, attributes, and data points.","A comprehensive technical explanation of the life and times of a young couple, from their marriage to their first marriage, and how they got married in the past, from the same era to today? - - - today - today.",NULL,0.81,0.3144,0.5622,300,300,1.0,352cdafa5458f753909e3cfc6a7e078c,images/2018/arxiv_0000234.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000235,Figure 235,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of the woman is sitting on a.,"A detailed layout showing a woman sitting on a train track with a suitcase and tattoo on her leg that shows key features, attributes, and data.","A comprehensive technical explanation of the art of travel, with a woman sitting on train tracks holding a suitcase and looking at it, in a field of grass and dirt, with trees and weeds, near railroad tracks, and a train tracks.",NULL,0.81,0.3559,0.583,300,300,1.0,e5f0d33ac70ce89c588d10c4654e1f6f,images/2018/arxiv_0000235.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000236,Figure 236,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a hospital bed on wheels.,"A detailed layout showing the bed, and the other parts of the hospital equipment in the room that shows key features, attributes, and data points in.","A comprehensive technical explanation of an emergency bed and its functions in the hospital room and how to use it for medical purposes? in the emergency room, hospital, this case, there is no other than for the patient, the bed.",NULL,0.81,0.3459,0.5779,300,300,1.0,dbcf3a41a43bed251bcab9c402e213b3,images/2018/arxiv_0000236.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000237,Figure 237,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 14: Lime-picked examples on the COCO test set.,A diagram of the four pictures show.,A detailed layout showing the different stages of a baseball game and how to use it for the next photo.,"A comprehensive technical explanation of the history of baseball in pictures from the beginning to the end of the 20th century, including the history and future of the game of the major league baseball teams that played in the game, including.",Figure 14: Lime-picked examples on the COCO test set.,0.7179,0.2818,0.4999,530,1244,0.426,035b814c1b10661a50e286fd325c373d,images/2018/arxiv_0000237.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000238,Figure 238,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 15: Lime-picked examples on the Out of Context (OOC) set.,A diagram of a person standing next to.,A detailed layout showing the woman in grey standing next to the refrigerator on the side of the road.,"A comprehensive technical explanation of refrigerators on the side of the road, including how to use them and where to put it in the freezer box and what to use it? - - - a refrigerator on the road - side.",Figure 15: Lime-picked examples on the Out of Context (OOC) set.,0.7143,0.3092,0.5118,300,300,1.0,69bbc9495b0003782a2162e8d62b69a6,images/2018/arxiv_0000238.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000239,Figure 239,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a woman holding a string.,A detailed layout showing a woman standing on the beach and flying a kite with a chair in the background.,"A comprehensive technical explanation of kite flying in the sky with a woman standing on the beach looking at it and holding a string to her neck with a chair sitting on the side of it, in the water, in front of the background.",NULL,0.725,0.3188,0.5219,300,300,1.0,ab48500cf69c10a71b769caf9727751b,images/2018/arxiv_0000239.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000240,Figure 240,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,NULL,A diagram of a cat sleeping in a bed.,"A detailed layout showing a cat inside of a bed in front of a golden crown on a red carpet that shows key features, attributes, and data points in.","A comprehensive technical explanation of the royal throne cats are sleeping on today in their own thrones, and they can sleep in it all day long, or night, at night all night, for a week long, that time.",NULL,0.81,0.3168,0.5634,300,300,1.0,839e58b704865d28d4ef1762f82404da,images/2018/arxiv_0000240.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000241,Figure 241,scientific_figure,Adversarial Semantic Alignment for Improved Image Captions,arXiv Research Authors,1805.00063v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Adversarial Semantic Alignment for Improved Image Captions. arXiv:1805.00063v3,Figure 15: Lime-picked examples on the Out of Context (OOC) set.,A diagram of a man standing on top of.,A detailed layout showing a man standing in front of a small airplane and a photo of another person.,"A comprehensive technical explanation of a prop plane that is flying in the sky with a man standing by it and looking at it from the side of the road, while another man stands on a boat, and another man is looking at the other side.",Figure 15: Lime-picked examples on the Out of Context (OOC) set.,0.7071,0.2602,0.4836,622,1259,0.494,52527007907840a22e397f8d85f22bf0,images/2018/arxiv_0000241.png,https://arxiv.org/pdf/1805.00063v3.pdf
arxiv_0000242,Figure 242,scientific_figure,Do Neural Network Cross-Modal Mappings Really Bridge Modalities?,arXiv Research Authors,1805.07616v2,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Do Neural Network Cross-Modal Mappings Really Bridge Modalities?. arXiv:1805.07616v2,"Figure 1: Effect of applying a mapping f to a (dis-
connected) manifold M with three hypothetical
classes (■, ▲and •).",A diagram of a curved object with a.,"A detailed layout showing a curved area with red and green cones on it and a green dot at the end that shows key features, attributes, and data.","A comprehensive technical explanation of a curve in 3d, including the following point of view of a curved object with a green dot on it ' s center and a red dot at the bottom of the corner at the center of the curve.","Figure 1: Effect of applying a mapping f to a (dis-
connected) manifold M with three hypothetical
classes (■, ▲and •).",0.81,0.2894,0.5497,525,329,1.596,567815ae687757899106ffc16d8c923d,images/2018/arxiv_0000242.png,https://arxiv.org/pdf/1805.07616v2.pdf
arxiv_0000243,Figure 243,scientific_figure,Do Neural Network Cross-Modal Mappings Really Bridge Modalities?,arXiv Research Authors,1805.07616v2,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Do Neural Network Cross-Modal Mappings Really Bridge Modalities?. arXiv:1805.07616v2,NULL,A diagram of a meshed area with a blue.,"A detailed layout showing the meshing and the two colored objects in the screen, with a blue background.","A comprehensive technical explanation of the two dimensional mesh model for a machine - controlled object, with a blue mesh covering it and a green mesh covering the red object with four dimensional mesh over the top of the other one side.",NULL,0.7214,0.3218,0.5216,434,306,1.418,09f1a62d3eb5640dc17fb3dd4db2fbba,images/2018/arxiv_0000243.png,https://arxiv.org/pdf/1805.07616v2.pdf
arxiv_0000244,Figure 244,scientific_figure,Do Neural Network Cross-Modal Mappings Really Bridge Modalities?,arXiv Research Authors,1805.07616v2,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Do Neural Network Cross-Modal Mappings Really Bridge Modalities?. arXiv:1805.07616v2,"Figure 1: Effect of applying a mapping f to a (dis-
connected) manifold M with three hypothetical
classes (■, ▲and •).",A diagram of different types of motion.,"A detailed layout showing the three different types of wave motion and its corresponding functions, including a speed limit and the direction of the.","A comprehensive technical explanation of the electric field that is in the space between two points of view and a point of view is shown below the line of view of view, with a distance on the object at the origin and the opposite end.","Figure 1: Effect of applying a mapping f to a (dis-
connected) manifold M with three hypothetical
classes (■, ▲and •).",0.75,0.3013,0.5256,517,412,1.255,103490262eecc7724cb0c85f31a10d8b,images/2018/arxiv_0000244.png,https://arxiv.org/pdf/1805.07616v2.pdf
arxiv_0000245,Figure 245,scientific_figure,Do Neural Network Cross-Modal Mappings Really Bridge Modalities?,arXiv Research Authors,1805.07616v2,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Do Neural Network Cross-Modal Mappings Really Bridge Modalities?. arXiv:1805.07616v2,"Figure 2: Learning a nn model in Wiki (left),
IAPR TC-12 (middle) and ImageNet (right).",A diagram of different plots showing.,"A detailed layout showing the different stages of the e - minox and e - mnoxx that shows key features, attributes, and data points in detail for.","A comprehensive technical explanation of the first and second phase of the h2n1k1k3k1fx / s1m0k2k1 - t1k5k1x1k7k representing all details, specifications, and configurations of the system components in full context with annotated elements.","Figure 2: Learning a nn model in Wiki (left),
IAPR TC-12 (middle) and ImageNet (right).",0.81,0.2934,0.5517,479,248,1.931,52626ae2199be48aad58b79974734afc,images/2018/arxiv_0000245.png,https://arxiv.org/pdf/1805.07616v2.pdf
arxiv_0000246,Figure 246,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 1. Many tasks exhibit many-to-many relationships between
inputs and outputs. Taking image captioning as an example, a
single image can be described with multiple captions (top) and
likewise a single caption can accurately describe multiple images
(bottom). In this work, we leverage these relationships in the data
to learn multi-modal output mappings from sparse annotations.",A diagram of the family is preparing.,A detailed layout showing a family eating together at a kitchen table and several other people sitting at a table.,"A comprehensive technical explanation of the kitchen and dining table for a family gathering together in a kitchen table, four people sit at a wooden dining table in a brick wall with other kitchen tables and a couple of wood chairs and dining.","Figure 1. Many tasks exhibit many-to-many relationships between
inputs and outputs. Taking image captioning as an example, a
single image can be described with multiple captions (top) and
likewise a single caption can accurately describe multiple images
(bottom). In this work, we leverage these relationships in the data
to learn multi-modal output mappings from sparse annotations.",0.81,0.305,0.5575,462,264,1.75,e339e75186047f569f279f2cd2be9731,images/2018/arxiv_0000246.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000247,Figure 247,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 2. Given a dataset of sparse annotations sampled from a true
multi-modal input-output mapping (left), our approach leverages a
learned similarity space to perform a soft-transfer of annotations
between semantically related inputs (right) – effectively recovering
the underlying multi-modal mapping from few samples.",A diagram of the different stages of a.,"A detailed layout showing the different types of data processing tools used for each of the following platforms, including.",A comprehensive technical explanation of the three - dimensional model of the system for using a time tracking system to track time and distance datas for each of objects in a particular time period of the process on a given network.,"Figure 2. Given a dataset of sparse annotations sampled from a true
multi-modal input-output mapping (left), our approach leverages a
learned similarity space to perform a soft-transfer of annotations
between semantically related inputs (right) – effectively recovering
the underlying multi-modal mapping from few samples.",0.75,0.2614,0.5057,828,413,2.005,2bbd17a39f5c2cd9333f602f7af78967,images/2018/arxiv_0000247.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000248,Figure 248,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 2. Given a dataset of sparse annotations sampled from a true
multi-modal input-output mapping (left), our approach leverages a
learned similarity space to perform a soft-transfer of annotations
between semantically related inputs (right) – effectively recovering
the underlying multi-modal mapping from few samples.",A diagram of various types of quantum.,"A detailed layout showing the various functions of a computer system and its components, including the cpu, processor.","A comprehensive technical explanation of the quantum system and its applications in physics, by j krig, m d and t s krigh, m g kr, p v n r v, m, m b representing all details, specifications, and configurations of the system components in full.","Figure 2. Given a dataset of sparse annotations sampled from a true
multi-modal input-output mapping (left), our approach leverages a
learned similarity space to perform a soft-transfer of annotations
between semantically related inputs (right) – effectively recovering
the underlying multi-modal mapping from few samples.",0.75,0.2099,0.4799,550,1321,0.416,b70e96fc50c113d4f4a1c5fcf6ec3ad5,images/2018/arxiv_0000248.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000249,Figure 249,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 3. An diagram of our approach for sequential prediction
task. Only the relevant segments (via a soft attention mechanism)
from the neighbors output are used. See Sec. 2.2 for details.",A diagram of a network with multiple.,A detailed layout showing the flow of the application process for a system that is not compatible to the application.,"A comprehensive technical explanation of the basic algorithm for the system of the same function, and how it works? - page 3 - p1 - 2 - p2 - p4 - png1 - p7 - p representing all details, specifications, and configurations of the system components in.","Figure 3. An diagram of our approach for sequential prediction
task. Only the relevant segments (via a soft attention mechanism)
from the neighbors output are used. See Sec. 2.2 for details.",0.75,0.2279,0.4889,720,405,1.778,73a95f349a710fc0728fadffefaa1fc7,images/2018/arxiv_0000249.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000250,Figure 250,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 3. An diagram of our approach for sequential prediction
task. Only the relevant segments (via a soft attention mechanism)
from the neighbors output are used. See Sec. 2.2 for details.",A diagram of a multimodal mapping.,"A detailed layout showing the functions of the learning multi - modal mapping system, and the steps to use it.","A comprehensive technical explanation of the multiple - dimensional model for learning multi - modal mappings and the multi - dimensional models in the system, with two diagrams below, the same set of the following versions of the model,.","Figure 3. An diagram of our approach for sequential prediction
task. Only the relevant segments (via a soft attention mechanism)
from the neighbors output are used. See Sec. 2.2 for details.",0.7429,0.3012,0.5221,552,432,1.278,028a03bd9918859e30f1152253b879b0,images/2018/arxiv_0000250.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000251,Figure 251,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples. with simple L2 regularization to show that our objective
goes beyond such simple regularization schemes. Method
# observed
Average Precision@k labels
@1
@5
@10 CE+L2
1
46.82
51.03
54.18
Ours
49.23
53.46
57.12 In these experiments, we find evidence that our method AWA 1. Induces smoothness in the conditional distributions.
Fig. 4 shows a setting where both CE and its L2 regular-
ized version obtain similar test losses but differ drasti-
cally in the label assignments compared to our objective.
Specifically, Fig. 4(b) shows the conditional probabil-
ity P(y = 1|x) for test points from a CE+L2 trained
model. Even for the L2 regularized model, there is sig-
nificant overfitting with some regions of the rightmost
Gaussian confidently predicted as class 1. In contrast,
our approach shown in Fig. 4(c) results in near uniform
probability between classes 1 and 2 within the cluster.
2. Acts as a regularizer. Since our objective enforces that
neighboring data points have similar output distributions,
over-fitting by making overly confident predictions is
strongly penalized. This is evidenced by the low KL-
divergence w.r.t. to underlying data distribution that our
objective achieves (see Fig. 4(d)).
3. Improves sample efficiency. As shown in Fig. 4(d),
even with fewer samples as compared to Maximum
Likelihood training (with and without L2 regularization),
our model is able to more accurately model the true data-
distribution as evidenced by the significantly lower KL
divergence w.r.t. to the data-generating model. CE+L2
20%
52.74
57.48
63.71
Ours
56.79
61.54
66.28 CE+L2
1
27.10
31.40
35.62
Ours
29.32
33.19
38.83 CUB CE+L2
20%
32.42
35.94
39.21
Ours
35.64
38.20
43.01",A diagram of a cluster of different.,A detailed layout showing the distribution of random data in a system that is not compatible to the data.,"A comprehensive technical explanation of the correlation of an example of a single - celled cell and a three - cell cell in a single celled system, with different size cells and two celled cells, each celled,.","Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples. with simple L2 regularization to show that our objective
goes beyond such simple regularization schemes. Method
# observed
Average Precision@k labels
@1
@5
@10 CE+L2
1
46.82
51.03
54.18
Ours
49.23
53.46
57.12 In these experiments, we find evidence that our method AWA 1. Induces smoothness in the conditional distributions.
Fig. 4 shows a setting where both CE and its L2 regular-
ized version obtain similar test losses but differ drasti-
cally in the label assignments compared to our objective.
Specifically, Fig. 4(b) shows the conditional probabil-
ity P(y = 1|x) for test points from a CE+L2 trained
model. Even for the L2 regularized model, there is sig-
nificant overfitting with some regions of the rightmost
Gaussian confidently predicted as class 1. In contrast,
our approach shown in Fig. 4(c) results in near uniform
probability between classes 1 and 2 within the cluster.
2. Acts as a regularizer. Since our objective enforces that
neighboring data points have similar output distributions,
over-fitting by making overly confident predictions is
strongly penalized. This is evidenced by the low KL-
divergence w.r.t. to underlying data distribution that our
objective achieves (see Fig. 4(d)).
3. Improves sample efficiency. As shown in Fig. 4(d),
even with fewer samples as compared to Maximum
Likelihood training (with and without L2 regularization),
our model is able to more accurately model the true data-
distribution as evidenced by the significantly lower KL
divergence w.r.t. to the data-generating model. CE+L2
20%
52.74
57.48
63.71
Ours
56.79
61.54
66.28 CE+L2
1
27.10
31.40
35.62
Ours
29.32
33.19
38.83 CUB CE+L2
20%
32.42
35.94
39.21
Ours
35.64
38.20
43.01",0.785,0.2852,0.5351,180,180,1.0,fc9821b1e786710b1ea3b9fc9edd0cfe,images/2018/arxiv_0000251.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000252,Figure 252,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples. with simple L2 regularization to show that our objective
goes beyond such simple regularization schemes. Method
# observed
Average Precision@k labels
@1
@5
@10 CE+L2
1
46.82
51.03
54.18
Ours
49.23
53.46
57.12 In these experiments, we find evidence that our method AWA 1. Induces smoothness in the conditional distributions.
Fig. 4 shows a setting where both CE and its L2 regular-
ized version obtain similar test losses but differ drasti-
cally in the label assignments compared to our objective.
Specifically, Fig. 4(b) shows the conditional probabil-
ity P(y = 1|x) for test points from a CE+L2 trained
model. Even for the L2 regularized model, there is sig-
nificant overfitting with some regions of the rightmost
Gaussian confidently predicted as class 1. In contrast,
our approach shown in Fig. 4(c) results in near uniform
probability between classes 1 and 2 within the cluster.
2. Acts as a regularizer. Since our objective enforces that
neighboring data points have similar output distributions,
over-fitting by making overly confident predictions is
strongly penalized. This is evidenced by the low KL-
divergence w.r.t. to underlying data distribution that our
objective achieves (see Fig. 4(d)).
3. Improves sample efficiency. As shown in Fig. 4(d),
even with fewer samples as compared to Maximum
Likelihood training (with and without L2 regularization),
our model is able to more accurately model the true data-
distribution as evidenced by the significantly lower KL
divergence w.r.t. to the data-generating model. CE+L2
20%
52.74
57.48
63.71
Ours
56.79
61.54
66.28 CE+L2
1
27.10
31.40
35.62
Ours
29.32
33.19
38.83 CUB CE+L2
20%
32.42
35.94
39.21
Ours
35.64
38.20
43.01",A diagram of the three colored dots.,"A detailed layout showing the different types of the data for a single - celled cell phone, with three separate cells.","A comprehensive technical explanation of the distribution of n - and n - r - spectras in a single celled polycline cell, using the same number of nd - t - and 3 - cell membranes representing all details, specifications, and configurations of the.","Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples. with simple L2 regularization to show that our objective
goes beyond such simple regularization schemes. Method
# observed
Average Precision@k labels
@1
@5
@10 CE+L2
1
46.82
51.03
54.18
Ours
49.23
53.46
57.12 In these experiments, we find evidence that our method AWA 1. Induces smoothness in the conditional distributions.
Fig. 4 shows a setting where both CE and its L2 regular-
ized version obtain similar test losses but differ drasti-
cally in the label assignments compared to our objective.
Specifically, Fig. 4(b) shows the conditional probabil-
ity P(y = 1|x) for test points from a CE+L2 trained
model. Even for the L2 regularized model, there is sig-
nificant overfitting with some regions of the rightmost
Gaussian confidently predicted as class 1. In contrast,
our approach shown in Fig. 4(c) results in near uniform
probability between classes 1 and 2 within the cluster.
2. Acts as a regularizer. Since our objective enforces that
neighboring data points have similar output distributions,
over-fitting by making overly confident predictions is
strongly penalized. This is evidenced by the low KL-
divergence w.r.t. to underlying data distribution that our
objective achieves (see Fig. 4(d)).
3. Improves sample efficiency. As shown in Fig. 4(d),
even with fewer samples as compared to Maximum
Likelihood training (with and without L2 regularization),
our model is able to more accurately model the true data-
distribution as evidenced by the significantly lower KL
divergence w.r.t. to the data-generating model. CE+L2
20%
52.74
57.48
63.71
Ours
56.79
61.54
66.28 CE+L2
1
27.10
31.40
35.62
Ours
29.32
33.19
38.83 CUB CE+L2
20%
32.42
35.94
39.21
Ours
35.64
38.20
43.01",0.75,0.246,0.498,180,180,1.0,9a1fe8e7cd9725ec1c4a3e6fe69f184b,images/2018/arxiv_0000252.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000253,Figure 253,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples. with simple L2 regularization to show that our objective
goes beyond such simple regularization schemes. Method
# observed
Average Precision@k labels
@1
@5
@10 CE+L2
1
46.82
51.03
54.18
Ours
49.23
53.46
57.12 In these experiments, we find evidence that our method AWA 1. Induces smoothness in the conditional distributions.
Fig. 4 shows a setting where both CE and its L2 regular-
ized version obtain similar test losses but differ drasti-
cally in the label assignments compared to our objective.
Specifically, Fig. 4(b) shows the conditional probabil-
ity P(y = 1|x) for test points from a CE+L2 trained
model. Even for the L2 regularized model, there is sig-
nificant overfitting with some regions of the rightmost
Gaussian confidently predicted as class 1. In contrast,
our approach shown in Fig. 4(c) results in near uniform
probability between classes 1 and 2 within the cluster.
2. Acts as a regularizer. Since our objective enforces that
neighboring data points have similar output distributions,
over-fitting by making overly confident predictions is
strongly penalized. This is evidenced by the low KL-
divergence w.r.t. to underlying data distribution that our
objective achieves (see Fig. 4(d)).
3. Improves sample efficiency. As shown in Fig. 4(d),
even with fewer samples as compared to Maximum
Likelihood training (with and without L2 regularization),
our model is able to more accurately model the true data-
distribution as evidenced by the significantly lower KL
divergence w.r.t. to the data-generating model. CE+L2
20%
52.74
57.48
63.71
Ours
56.79
61.54
66.28 CE+L2
1
27.10
31.40
35.62
Ours
29.32
33.19
38.83 CUB CE+L2
20%
32.42
35.94
39.21
Ours
35.64
38.20
43.01",A diagram of the plot shows how the.,A detailed layout showing the different levels of the cee loss and the corresponding loss of each loss.,"A comprehensive technical explanation of the cel loss and loss curve for each two years of use, from the eu / us data center on the web site, 2007 - edg, eu / eu / eci, eu representing all details, specifications, and configurations of the system.","Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples. with simple L2 regularization to show that our objective
goes beyond such simple regularization schemes. Method
# observed
Average Precision@k labels
@1
@5
@10 CE+L2
1
46.82
51.03
54.18
Ours
49.23
53.46
57.12 In these experiments, we find evidence that our method AWA 1. Induces smoothness in the conditional distributions.
Fig. 4 shows a setting where both CE and its L2 regular-
ized version obtain similar test losses but differ drasti-
cally in the label assignments compared to our objective.
Specifically, Fig. 4(b) shows the conditional probabil-
ity P(y = 1|x) for test points from a CE+L2 trained
model. Even for the L2 regularized model, there is sig-
nificant overfitting with some regions of the rightmost
Gaussian confidently predicted as class 1. In contrast,
our approach shown in Fig. 4(c) results in near uniform
probability between classes 1 and 2 within the cluster.
2. Acts as a regularizer. Since our objective enforces that
neighboring data points have similar output distributions,
over-fitting by making overly confident predictions is
strongly penalized. This is evidenced by the low KL-
divergence w.r.t. to underlying data distribution that our
objective achieves (see Fig. 4(d)).
3. Improves sample efficiency. As shown in Fig. 4(d),
even with fewer samples as compared to Maximum
Likelihood training (with and without L2 regularization),
our model is able to more accurately model the true data-
distribution as evidenced by the significantly lower KL
divergence w.r.t. to the data-generating model. CE+L2
20%
52.74
57.48
63.71
Ours
56.79
61.54
66.28 CE+L2
1
27.10
31.40
35.62
Ours
29.32
33.19
38.83 CUB CE+L2
20%
32.42
35.94
39.21
Ours
35.64
38.20
43.01",0.7779,0.2977,0.5378,804,506,1.589,3639fdf65f151915bded41b540a4f0b6,images/2018/arxiv_0000253.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000254,Figure 254,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples.",A diagram of the different types of.,A detailed layout showing the results of a multicolored particle experiment on a white background with different data.,"A comprehensive technical explanation of the two - dimensional model of a cluster of nanoclot structures, including the following model and the corresponding model of the same model of each component of the different materials in the system,.","Figure 4. (a) Our toy experiment uses synthetic data with uniform label mixing within each cluster. (b) We find that cross-entropy (CE)
training (even with L2 regularization) results in overfitting – for instance, some regions of the rightmost cluster are very confidently
predicted as class 1 despite the true distribution being unbiased. (c) In contrast, our approach accurately predicts equal likelihood within
clusters. (d) Compared to training with CE loss, our approach accurately matches the true distribution as seen by the significantly lower
KL-divergence values w.r.t. the true-distribution while utilizing an order of magnitude fewer samples.",0.81,0.2925,0.5513,1224,424,2.887,7758768cab96d428c078ec6e52caa640,images/2018/arxiv_0000254.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000255,Figure 255,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 5. Our method is sample efficient compared to standard
cross-entropy training. On the Flickr8k image-captioning task, our
method performs comparably to CE with the full dataset while
only using ∼3K images (left). Similar trends exist as the number
of annotations per image is increased (right). in Fig. 5 (left), our approach obtains a Recall5@100 score
of 1.58 using only 5K images which is very close to what is
obtained using all 8k images (=1.60) in the first case. In the
second case, we observe in Fig. 5 (right) that using only 3
captions per image (1.65) nearly obtains the same retrieval
score as using all the 5 captions (=1.66). This demonstrates
that our approach can lead to efficient learning of the true
distribution even in data-sparse regimes. While the primary focus of our work is to learn multi-modal
mappings even with access to uni-modal datasets, we
observe that our proposed method performs competitively
when trained and evaluated under standard captioning
settings i.e. all 5 captions are used during training and one
best output is evaluated (as against using oracle metrics).
For instance, our method achieves a METEOR and CIDEr
score of 0.28 and 1.14 respectively, slightly outperforming
Lu et al. (2017) (0.27 and 1.09) on the COCO-captioning
task. We observe similar trends on question-generation and
include detailed results in the supplement. 6. Conclusion In this work, we propose a novel objective that incorpo-
rates the inductive bias that the outputs of neighboring data
points can be used to provide additional supervision espe-
cially when obtaining exhaustive annotations is expensive
or worse, intractable. The proposed objective allows the
model to place beliefs on multiple plausible outputs while
still observing only one annotation per input. We first study
the properties of our method on a synthetic dataset where the
underlying data-distribution is known allowing us to control
the difficulty of the experiments and directly evaluate the
learnt posteriors. Further, we replicate this toy setting on
a real-world multi-label prediction problem using standard
attribute datasets. Finally, we show that our approach leads
to better quality outputs with higher diversity on two well-
established visually grounded language-generation tasks –
captioning and question generation. We observe that our
approach outperforms various ablations and baselines on
both tasks on the various evaluation metrics used.",A diagram of the average and depicting.,A detailed layout showing the relative and relative range of the micehie mouses during the first phase.,"A comprehensive technical explanation of the detection of the anti - active body - specific gene from the human and non - active gene cells in the mouses of mice, including the mouse mouses and mouses, mouses - mouse.","Figure 5. Our method is sample efficient compared to standard
cross-entropy training. On the Flickr8k image-captioning task, our
method performs comparably to CE with the full dataset while
only using ∼3K images (left). Similar trends exist as the number
of annotations per image is increased (right). in Fig. 5 (left), our approach obtains a Recall5@100 score
of 1.58 using only 5K images which is very close to what is
obtained using all 8k images (=1.60) in the first case. In the
second case, we observe in Fig. 5 (right) that using only 3
captions per image (1.65) nearly obtains the same retrieval
score as using all the 5 captions (=1.66). This demonstrates
that our approach can lead to efficient learning of the true
distribution even in data-sparse regimes. While the primary focus of our work is to learn multi-modal
mappings even with access to uni-modal datasets, we
observe that our proposed method performs competitively
when trained and evaluated under standard captioning
settings i.e. all 5 captions are used during training and one
best output is evaluated (as against using oracle metrics).
For instance, our method achieves a METEOR and CIDEr
score of 0.28 and 1.14 respectively, slightly outperforming
Lu et al. (2017) (0.27 and 1.09) on the COCO-captioning
task. We observe similar trends on question-generation and
include detailed results in the supplement. 6. Conclusion In this work, we propose a novel objective that incorpo-
rates the inductive bias that the outputs of neighboring data
points can be used to provide additional supervision espe-
cially when obtaining exhaustive annotations is expensive
or worse, intractable. The proposed objective allows the
model to place beliefs on multiple plausible outputs while
still observing only one annotation per input. We first study
the properties of our method on a synthetic dataset where the
underlying data-distribution is known allowing us to control
the difficulty of the experiments and directly evaluate the
learnt posteriors. Further, we replicate this toy setting on
a real-world multi-label prediction problem using standard
attribute datasets. Finally, we show that our approach leads
to better quality outputs with higher diversity on two well-
established visually grounded language-generation tasks –
captioning and question generation. We observe that our
approach outperforms various ablations and baselines on
both tasks on the various evaluation metrics used.",0.7179,0.2306,0.4743,432,288,1.5,16c9f4a52d01c9919d9d04fdb2e5acc9,images/2018/arxiv_0000255.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000256,Figure 256,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,NULL,A diagram of the line graph of the.,"A detailed layout showing the results of a training cycle in a single - cell phone system, with a line graphing the number of active cells.","A comprehensive technical explanation of the average training time for the training area for the first course of the course, a / e, and a / l / e / e - 1 / g / f, in - b representing all details, specifications, and configurations of the system.",NULL,0.93,0.3109,0.6205,432,288,1.5,e1d3b260119974c3ddfd42ecfe590e11,images/2018/arxiv_0000256.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000257,Figure 257,scientific_figure,Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations,arXiv Research Authors,1806.02934v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learn from Your Neighbor: Learning Multi-modal Mappings from Sparse Annotations. arXiv:1806.02934v1,"Figure 5. Our method is sample efficient compared to standard
cross-entropy training. On the Flickr8k image-captioning task, our
method performs comparably to CE with the full dataset while
only using ∼3K images (left). Similar trends exist as the number
of annotations per image is increased (right).",A diagram of graphs from spare motions.,A detailed layout showing the differences between two graphs and the same graphing process in a computer application.,"A comprehensive technical explanation of graphs from spare motions and their applications, including the speed curves and the speed of the motor cycle, which is the same speed as shown in the curve as shown below the curve??? -.","Figure 5. Our method is sample efficient compared to standard
cross-entropy training. On the Flickr8k image-captioning task, our
method performs comparably to CE with the full dataset while
only using ∼3K images (left). Similar trends exist as the number
of annotations per image is increased (right).",0.81,0.2815,0.5457,551,313,1.76,d751dc192aa548b35edc06a8e3eaeb98,images/2018/arxiv_0000257.png,https://arxiv.org/pdf/1806.02934v1.pdf
arxiv_0000258,Figure 258,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 1. The input of the generalized Earley parser is a matrix of
probabilities of each label for each frame, given by an arbitrary
classifier. The parser segments and labels the sequence data into
a label sentence in the language of a given grammar. Future
predictions are then made based on the grammar.",A diagram of the process for using a.,"A detailed layout showing the process of a digital image processing process, including a single layer of photos.","A comprehensive technical explanation of the algorithm for the first - generation of optically - controlled data processing system, based on the image of a computer screen and a computer monitor screen or tablet screen, in a computer, with multiple.","Figure 1. The input of the generalized Earley parser is a matrix of
probabilities of each label for each frame, given by an arbitrary
classifier. The parser segments and labels the sequence data into
a label sentence in the language of a given grammar. Future
predictions are then made based on the grammar.",0.81,0.2593,0.5347,441,308,1.432,337a51f4ecaa232556102735e9a563f4,images/2018/arxiv_0000258.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000259,Figure 259,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,Figure 2. An example of the original Earley parser.,A diagram of the sequence of an array.,"A detailed layout showing the basic rules for using a key parser in python and python programmings that shows key features, attributes, and data.","A comprehensive technical explanation of the key parser with examples for each of its functions and functions, including a number of key parses to the main system, the first one in the key, and second, the third,.",Figure 2. An example of the original Earley parser.,0.81,0.3275,0.5688,448,540,0.83,b98d73a1e34be0af4577b11b425518ad,images/2018/arxiv_0000259.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000260,Figure 260,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 3. Prefix search according to grammar. A classifier is ap-
plied to a 5-frame signal and outputs a probability matrix (bot-
tom right) as the input to our algorithm. The proposed algorithm
expands a grammar prefix tree (left), where “e” represents termi-
nation. It finally outputs the best label “0 + 1” with probability
0.43. The probabilities of children nodes do not sum to 1 since the
grammatically incorrect nodes are eliminated from the search.",A diagram of a tree with three numbers.,"A detailed layout showing the graphing tree for the first three trees of the series, which is also an example of a tree with a number of different.","A comprehensive technical explanation of the tree model for the data visual model for trees, with two rows of numbers in each row and one column in the same row, and the number of rows in the middle row on the same.","Figure 3. Prefix search according to grammar. A classifier is ap-
plied to a 5-frame signal and outputs a probability matrix (bot-
tom right) as the input to our algorithm. The proposed algorithm
expands a grammar prefix tree (left), where “e” represents termi-
nation. It finally outputs the best label “0 + 1” with probability
0.43. The probabilities of children nodes do not sum to 1 since the
grammatically incorrect nodes are eliminated from the search.",0.81,0.3019,0.556,208,265,0.785,225213d97eeea81c76ac35cc358fde94,images/2018/arxiv_0000260.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000261,Figure 261,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 4. Qualitative results of segmentation results. In each group
of four segmentations, the rows from the top to the bottom show
the results of: 1) ground-truth, 2) ST-AOG + Earley, 3) Bi-LSTM,
and 4) Bi-LSTM + generalized Earley parser. The results show
(a) corrections and (b) insertions by our algorithm on the initial
segment-wise labels given by the classifier (Bi-LSTM). for detection, but significantly outperforms the classifier for
predictions. The modifications on classifier outputs (correc-
tions and insertions in Figure 4) are minor but important to
make the sentences grammatically correct, thus high-quality
predictions can be made. How useful is the grammar for activity modeling? From
Table 2, Table 3, Table 5 and Table 6 we can see that both ST-
AOG and generalized Earley parser outperforms Bi-LSTM
for prediction. Prediction algorithms need to give different
outputs for similar inputs based on the observation history.
Hence the non-Markovian property of grammars is useful
for activity modeling, especially for future prediction. How robust is the generalized Earley parser? Compar-
ing Table 3 and Table 6 we can see that there is a perfor-
mance drop when the action sequences are more unpre-
dictable (in the Watch-n-Patch dataset). But it is capable of
improving over the noisy classifier inputs and significantly
outperforms the other methods. It is also robust in the sense
that it can always find the best sentence in a given language
that best explains the classifier outputs. 7. Conclusions We proposed a generalized Earley parser for parsing se-
quence data according to symbolic grammars. Detections
and predictions are made by the parser given the probabilis-
tic outputs from any classifier. We are optimistic about and
interested in further applications of the generalized Earley
parser. In general, we believe this is a step towards the goal
of integrating the connectionist and symbolic approaches.",A diagram of the various colors of a.,"A detailed layout showing the colors of the entire set of color bars in various shades of red, blue and green.","A comprehensive technical explanation of the internet age of people in the past 50 years to the present, according with the number of users at each age of the device, by age of their devices and age of users, as well.","Figure 4. Qualitative results of segmentation results. In each group
of four segmentations, the rows from the top to the bottom show
the results of: 1) ground-truth, 2) ST-AOG + Earley, 3) Bi-LSTM,
and 4) Bi-LSTM + generalized Earley parser. The results show
(a) corrections and (b) insertions by our algorithm on the initial
segment-wise labels given by the classifier (Bi-LSTM). for detection, but significantly outperforms the classifier for
predictions. The modifications on classifier outputs (correc-
tions and insertions in Figure 4) are minor but important to
make the sentences grammatically correct, thus high-quality
predictions can be made. How useful is the grammar for activity modeling? From
Table 2, Table 3, Table 5 and Table 6 we can see that both ST-
AOG and generalized Earley parser outperforms Bi-LSTM
for prediction. Prediction algorithms need to give different
outputs for similar inputs based on the observation history.
Hence the non-Markovian property of grammars is useful
for activity modeling, especially for future prediction. How robust is the generalized Earley parser? Compar-
ing Table 3 and Table 6 we can see that there is a perfor-
mance drop when the action sequences are more unpre-
dictable (in the Watch-n-Patch dataset). But it is capable of
improving over the noisy classifier inputs and significantly
outperforms the other methods. It is also robust in the sense
that it can always find the best sentence in a given language
that best explains the classifier outputs. 7. Conclusions We proposed a generalized Earley parser for parsing se-
quence data according to symbolic grammars. Detections
and predictions are made by the parser given the probabilis-
tic outputs from any classifier. We are optimistic about and
interested in further applications of the generalized Earley
parser. In general, we believe this is a step towards the goal
of integrating the connectionist and symbolic approaches.",0.7429,0.2849,0.5139,2800,500,5.6,a94d7c172b60c6e339f56c1746920039,images/2018/arxiv_0000261.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000262,Figure 262,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 4. Qualitative results of segmentation results. In each group
of four segmentations, the rows from the top to the bottom show
the results of: 1) ground-truth, 2) ST-AOG + Earley, 3) Bi-LSTM,
and 4) Bi-LSTM + generalized Earley parser. The results show
(a) corrections and (b) insertions by our algorithm on the initial
segment-wise labels given by the classifier (Bi-LSTM).",A diagram of the different types of.,A detailed layout showing the different color combinations of a colored ray parser and the different colors of the ray.,"A comprehensive technical explanation of the color coding for a color - coded ray parser, and its four primary colors, are shown in this diagram from a to b to c - b to dg - 3 / s / w representing all details, specifications, and configurations of.","Figure 4. Qualitative results of segmentation results. In each group
of four segmentations, the rows from the top to the bottom show
the results of: 1) ground-truth, 2) ST-AOG + Earley, 3) Bi-LSTM,
and 4) Bi-LSTM + generalized Earley parser. The results show
(a) corrections and (b) insertions by our algorithm on the initial
segment-wise labels given by the classifier (Bi-LSTM).",0.75,0.3245,0.5373,553,547,1.011,61e6011b57b599dd3a378f6db7d2e7f9,images/2018/arxiv_0000262.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000263,Figure 263,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,Figure 1. Production rules of an example English grammar. “|”,A diagram of some type of grammars for.,A detailed layout showing the basic steps to make a formal grammar and its structure and meaning in english.,"A comprehensive technical explanation of the grammar and its uses in english language, including examples of using thesauruss to use to make thesaurussauruss and their own words in english ornamel ornames as.",Figure 1. Production rules of an example English grammar. “|”,0.7357,0.3196,0.5276,554,954,0.581,3f65f6e8a4e8bc2a22a75491299a54e7,images/2018/arxiv_0000263.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000264,Figure 264,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,Figure 2. Parsing a sentence according to grammar productions.,"A diagram of a sentence, with the.","A detailed layout showing the structure of an object in the text book, including a sentence and an object.","A comprehensive technical explanation of the structure of an object in a sentence or sentence, this is a subject verb object object noun noun noun sentence this is subject subject subject verb verb object this is an object this ' s subject this.",Figure 2. Parsing a sentence according to grammar productions.,0.7286,0.3116,0.5201,491,164,2.994,7b7ae312ae4c61dd9b388b7e070e0448,images/2018/arxiv_0000264.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000265,Figure 265,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,Figure 3. An example of the generalized Earley parser. This exam-,A diagram of a computer system with an.,"A detailed layout showing the electronic components for a computer system, including a processor and a printer,.","A comprehensive technical explanation of the general and operational system for a solar generator and its operation in the industrial sector, using the general series of circuitry systems of the solars and the system at the same time on the circuit.",Figure 3. An example of the generalized Earley parser. This exam-,0.75,0.2062,0.4781,554,1374,0.403,b9a9f633f113ee0c52b5e49945225092,images/2018/arxiv_0000265.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000266,Figure 266,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,Figure 4. Visualization of the selected super-pixels for extracting kernel descriptors,A diagram of the various images of.,"A detailed layout showing how the computer workspace is used for all of the work on the desk that shows key features, attributes, and data points in.","A comprehensive technical explanation of a computer system with multiple images of the same device, including a monitor and keyboard, for each of its users to use on the same level of the computer devices, with the same purpose,.",Figure 4. Visualization of the selected super-pixels for extracting kernel descriptors,0.81,0.2345,0.5222,680,640,1.062,e85b24869df20b213038128c95128bc5,images/2018/arxiv_0000266.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000267,Figure 267,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,Generalized Earley Parser,A diagram of the percentage of people.,A detailed layout showing the number of different colored bars in each of the rows of each color scheme.,"A comprehensive technical explanation of the top 10 countries for data analytics and web design, including the u s, canada, and australia, in the year 1999 - 2012 and 2011 - 2013, according with over 3, the same.",Generalized Earley Parser,0.7214,0.2709,0.4961,716,128,5.594,0d12bb56ccb418ad8ecda820f7811e25,images/2018/arxiv_0000267.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000268,Figure 268,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the number of people who.,"A detailed layout showing the different colors of the color of the rainbows and their meanings, including the red.","A comprehensive technical explanation of the different color shades in the lines on this panel are blue, green and red, and are white and red and blue, with small blue and orange, and yellow and red stripes on the same.",NULL,0.75,0.2213,0.4857,716,128,5.594,0b7974dde03977e029d44131f97edd2d,images/2018/arxiv_0000268.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000269,Figure 269,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the percentage of people.,"A detailed layout showing the number of people in each country with different nationalitiess and nationalitys, including.","A comprehensive technical explanation of the u s national debt index for 2013 and beyond, including the percentage of debts sold by country in the us dollars per share price of goods, from each country in 2013 to the united states.",NULL,0.75,0.2417,0.4959,716,128,5.594,50a2f421de5b093bbb52975fb88c0aa9,images/2018/arxiv_0000269.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000270,Figure 270,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the percentage of people.,A detailed layout showing the different color palettes on the same page in each section of the page.,"A comprehensive technical explanation of the different color schemes used for each type of product, from the top to bottom, including blue, red, and green, and blue, and black, and white, in this color scheme of the same.",NULL,0.7071,0.2529,0.48,716,128,5.594,8be246bb7478751acd0a2bea68959487,images/2018/arxiv_0000270.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000271,Figure 271,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the colors of different.,"A detailed layout showing the different color combinations of the lines in this image, as well as the corresponding colors.","A comprehensive technical explanation of the color code for color codes in web design, including the following colors and their meanings and examples of color codes - based on the original color code codes and symbols - coding - stock -.",NULL,0.75,0.2897,0.5199,716,128,5.594,79dd4c121f0cc73788177321756568a2,images/2018/arxiv_0000271.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000272,Figure 272,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the different colors of a.,"A detailed layout showing the different colors of the rainbow bars, including red, blue and green that shows key features, attributes, and data.","A comprehensive technical explanation of the color code for the four - color color scheme of the computer system, including the colors in purple and red and blue, and blue and green and black, for the red, with a third.",NULL,0.81,0.2405,0.5252,716,128,5.594,1c3426331a4c7ed6360a8ee09a18a29f,images/2018/arxiv_0000272.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000273,Figure 273,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the colors of different.,"A detailed layout showing the different colors of the lines in this chart, and how to use them that shows key features, attributes, and data points.","A comprehensive technical explanation of the different color shades of the same color scheme on the line chart of the colors in each panel, as described by the following one color scheme of the color palettes in the line of the lines.",NULL,0.87,0.2505,0.5603,716,128,5.594,11b539de71348de81a27a9e8fa2722f2,images/2018/arxiv_0000273.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000274,Figure 274,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the color bars in this.,"A detailed layout showing the different colors of the blue and red stripes in each row, as well as that shows key features, attributes, and data.","A comprehensive technical explanation of the color palette in adobe indesignate your design in seconds with this handy guide to color palettes, you can create a complete set up any color scheme of lines and choose from one color.",NULL,0.81,0.2382,0.5241,716,128,5.594,e661040b7700418922004b34c9bba27c,images/2018/arxiv_0000274.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000275,Figure 275,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",A diagram of different colors of blue.,"A detailed layout showing the color of the lines in different shades of blue, red and green, with white.","A comprehensive technical explanation of the color scheme for the colors in this chart are blue, red and green, and black and white and with a few of a hint of a bit of a third color on it is a hint representing all details, specifications, and.","Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",0.7214,0.2536,0.4875,716,128,5.594,89e8dc689535590328410e8a6eb45507,images/2018/arxiv_0000275.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000276,Figure 276,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",A diagram of the number of people in.,"A detailed layout showing the different color combinations of a bar chart in one place, and the other half.","A comprehensive technical explanation of the different types of protein in a food processor for the production of proteins and other foods, including poultry, beef, fish, fish and vegetables, and rice, and meats, and fish.","Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",0.7921,0.2664,0.5292,716,128,5.594,7490713d54c932e36f35f2272d61fdf8,images/2018/arxiv_0000276.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000277,Figure 277,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",A diagram of the various colors of the.,"A detailed layout showing the colors of the flag of the united states, with different horizontals and verticals.","A comprehensive technical explanation of the percentage of people with a particular disability in the united states, by age and race, 2006 - 2012, based on the u s / s / t / n, t / t, 2012 representing all details, specifications, and.","Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",0.75,0.2333,0.4917,716,128,5.594,f473a9af46866f598c0e92c12b9b73cf,images/2018/arxiv_0000277.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000278,Figure 278,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",A diagram of the four different colors.,"A detailed layout showing the color of the spectrums of red, green and blue on a white background that shows key features, attributes, and data.","A comprehensive technical explanation of the different color codes for each of the four types of colors in the computer screen, including red, green and blue, and yellow, from the white to the bottom - right to the top - left.","Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1) ground-truth, 2) results of ST-AOG, 3) Bi-LSTM, and 4) Bi-LSTM + generalized Earley parser.",0.81,0.2731,0.5415,716,128,5.594,9ee824826916fcd38860a87f8f00f375,images/2018/arxiv_0000278.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000279,Figure 279,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the percentage of people.,A detailed layout showing the different colors of the flag of france and the country ' s national flag.,"A comprehensive technical explanation of the performance of the cme - based software for business and product management in the market environment, including the performance score of the company ' s performance ratings and performance of its.",NULL,0.7179,0.232,0.4749,716,128,5.594,876d7e1343633b78dbe1f9110b49313f,images/2018/arxiv_0000279.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000280,Figure 280,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of different colors of.,"A detailed layout showing the different color palettes used for each of the colors in this image, including red.","A comprehensive technical explanation of the different color schemes used in a computer application, including the green and red stripes on the bottom of the screen are horizontal lines in the same direction of the line with the green stripe at the.",NULL,0.75,0.2353,0.4927,716,128,5.594,a7a9283bbe3a1e3c40bb71d8fdff91a7,images/2018/arxiv_0000280.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000281,Figure 281,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the different colors of a.,"A detailed layout showing the color of the rainbow chart, with different colors and sizes of each color.","A comprehensive technical explanation of color theory in the design process and the use of colors to create a line chart of color schemes, including green, orange, yellow, red, purple, red and blue, green, and orange.",NULL,0.7814,0.253,0.5172,716,128,5.594,f3bc78d212e1c42f92ab293dc4e18b14,images/2018/arxiv_0000281.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000282,Figure 282,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the various colors of the.,"A detailed layout showing the different colors of a fruit and vegetable plant, including red, yellow, green, orange, and red.","A comprehensive technical explanation of the different color - changing temperatures in a home kitchen, including temperature and humidity levels, according with thermometer and temperature scale, in red and yellow, green, orange, yellow, red.",NULL,0.81,0.2412,0.5256,716,128,5.594,13d7de283eaa282185a7fd20aeee661e,images/2018/arxiv_0000282.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000283,Figure 283,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the color bars in.,"A detailed layout showing the various color schemes for each of the squares in this chart, each with different colors.","A comprehensive technical explanation of the color scale on a computer screen, with a few different colors and sizes to choose from, including red, green and yellow, orange and red, and yellow and purple, and red photo - stock photo.",NULL,0.81,0.2459,0.528,716,128,5.594,ab775797700a40f091a12517a2f1b935,images/2018/arxiv_0000283.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000284,Figure 284,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the different colors in.,"A detailed layout showing the different colors of the rainbow - colored bars of color, from red to yellow.","A comprehensive technical explanation of color codes for the colors of the rainbows, and how they are used to make them stand out in any color scheme of the world ornameblearchnice color scheme, or any type.",NULL,0.7286,0.2919,0.5102,716,128,5.594,445695e511af5d5e13042bdf76ce778d,images/2018/arxiv_0000284.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000285,Figure 285,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the different colors of.,"A detailed layout showing the different colors of the rainbow bars in each color scheme, and a white background.","A comprehensive technical explanation of the different color combinations in the samples of the sample samples for each tester, and their corresponding colors to which one is different color scheme of the number of the testerbercented colors.",NULL,0.75,0.2971,0.5235,716,128,5.594,93b8a2a144f983dcf2f12ef7d0a3a258,images/2018/arxiv_0000285.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000286,Figure 286,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of different colors of.,"A detailed layout showing the color of the green and red bars in this image, with a white background.","A comprehensive technical explanation of the green color palette in a computer application, with a simple explanation of what colors are used for each color scheme in this image, and how they are used in the green, green color schemes,.",NULL,0.7107,0.2908,0.5008,716,128,5.594,c4c12481515fc33ab79fa27646c87ace,images/2018/arxiv_0000286.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000287,Figure 287,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the different colors of a.,"A detailed layout showing the different colors of green and red stripes of the same color scheme, with a white.","A comprehensive technical explanation of color theory in the design of a flat paneled design, with the colors red, green and blue on each panel, and white background is a horizontal line of a horizontal pattern that shows a horizontal.",NULL,0.7464,0.2715,0.509,716,128,5.594,0b84e4b62fe7e47c6362f69cad3991b7,images/2018/arxiv_0000287.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000288,Figure 288,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the percentage of people.,"A detailed layout showing the different color schemes of a bar chart, with the red and green bars that shows key features, attributes, and data.","A comprehensive technical explanation of the different types of software performance indicators in the web development process, including performance ratings and performance rating score for each product or service level of the company or product.",NULL,0.87,0.2403,0.5552,716,128,5.594,3715d2053bda6eae7bf7621c9cf1d102,images/2018/arxiv_0000288.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000289,Figure 289,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of the color scheme for the.,A detailed layout showing the different colors of the rainbows on a white background stock photo - 5397.,"A comprehensive technical explanation of the color scale for each of the three color bars in this chart is also useful to see if you have a specific color and a specific set of colors, how they are used in them??.",NULL,0.7214,0.2466,0.484,716,128,5.594,e6fe31ed9e0c9857e4d4bd16a57e9547,images/2018/arxiv_0000289.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000290,Figure 290,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,NULL,A diagram of thermometers and.,A detailed layout showing the different colors of a temperature chart for each of the four days in the day.,A comprehensive technical explanation of a temperature indicator for thermometers - screenshote com - temperature indicator - temperature gauge - chart - temperature - temperature png - 3 png png jpg - png com.,NULL,0.7921,0.2702,0.5312,716,128,5.594,2ce11d267f6978940364b4558d5767bd,images/2018/arxiv_0000290.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000291,Figure 291,scientific_figure,Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction,arXiv Research Authors,1806.03497v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generalized Earley Parser: Bridging Symbolic Grammars and Sequence Data for Future Prediction. arXiv:1806.03497v1,"Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1)",A diagram of the various colors of.,A detailed layout showing the multiple layers of the color scheme for a multicolored pattern of lines.,"A comprehensive technical explanation of the color coded data visual for the computer and data engineering library, featuring color coded lines of data in different colors, including the same size and fonts and numbers, from the same hues,.","Figure 5. Qualitative results of segmentation results. In each group of four segmentations, the rows from the top to the bottom shows: 1)",0.7743,0.3045,0.5394,1224,1354,0.904,0925c9a73e15440221339c2621147f4c,images/2018/arxiv_0000291.png,https://arxiv.org/pdf/1806.03497v1.pdf
arxiv_0000292,Figure 292,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure 1: ACD illustrated through the toy example of predicting the phrase “not very good” as
negative. Given the network and prediction, ACD constructs a hierarchy of meaningful phrases and
provides importance scores for each identified phrase. In this example, ACD identifies that “very”
modifies “good” to become the very positive phrase “very good”, which is subsequently negated by
”not” to produce the negative phrase “not very good”. Best viewed in color.",A diagram of the two types of dmn and.,"A detailed layout showing the two main types of dn and dn predictions, including negative, not very good, very good.","A comprehensive technical explanation of dnv prediction and interpretation in python and c + + + dn = c + dm + dy = dn i dn negative negative negative not very good very good negative not representing all details, specifications, and configurations.","Figure 1: ACD illustrated through the toy example of predicting the phrase “not very good” as
negative. Given the network and prediction, ACD constructs a hierarchy of meaningful phrases and
provides importance scores for each identified phrase. In this example, ACD identifies that “very”
modifies “good” to become the very positive phrase “very good”, which is subsequently negated by
”not” to produce the negative phrase “not very good”. Best viewed in color.",0.75,0.3059,0.528,715,344,2.078,33a2bfb4463730a5069e0cb694f42ba9,images/2018/arxiv_0000292.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000293,Figure 293,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure 2: ACD interpretation of an LSTM predicting sentiment. Blue is positive sentiment, white
is neutral, red is negative. The bottom row displays CD scores for individual words in the sentence.
Higher rows display important phrases identified by ACD, along with their CD scores, converging
to the model’s (incorrect) prediction in the top row. (Best viewed in color)",A diagram of the different types of.,"A detailed layout showing the different types of online dating sites in each country, from one to five.","A comprehensive technical explanation of the internet age system, including the internet information and the internet address to the user and the user area for each of the devices, in the user, as well as well, as described.","Figure 2: ACD interpretation of an LSTM predicting sentiment. Blue is positive sentiment, white
is neutral, red is negative. The bottom row displays CD scores for individual words in the sentence.
Higher rows display important phrases identified by ACD, along with their CD scores, converging
to the model’s (incorrect) prediction in the top row. (Best viewed in color)",0.7179,0.193,0.4555,2702,794,3.403,17cab705509af3cce69277984c4e3f63,images/2018/arxiv_0000293.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000294,Figure 294,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure 2: ACD interpretation of an LSTM predicting sentiment. Blue is positive sentiment, white
is neutral, red is negative. The bottom row displays CD scores for individual words in the sentence.
Higher rows display important phrases identified by ACD, along with their CD scores, converging
to the model’s (incorrect) prediction in the top row. (Best viewed in color)",A diagram of a number of different.,A detailed layout showing the number of people in each area of the country and the number on each side.,"A comprehensive technical explanation of the astma and its applications in the ast ' s computer environment, for the use of the computer network and the asts and ast s computers, astmast ' t representing all details, specifications, and.","Figure 2: ACD interpretation of an LSTM predicting sentiment. Blue is positive sentiment, white
is neutral, red is negative. The bottom row displays CD scores for individual words in the sentence.
Higher rows display important phrases identified by ACD, along with their CD scores, converging
to the model’s (incorrect) prediction in the top row. (Best viewed in color)",0.7179,0.1852,0.4516,1224,919,1.332,6e926ed3c78d5036655dd11a61f44c1a,images/2018/arxiv_0000294.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000295,Figure 295,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure 3: ACD interpretation for a VGG network prediction, described in 4.2.1. ACD shows that
the CNN is focusing on skates to predict the class “puck”, indicating that the model has captured
dataset bias. The top row shows the original image, logits for the five top-predicted classes, and the
CD superpixel-level scores for those classes. The second row shows separate image patches ACD
has identified as being independently predictive of the class “puck”. Starting from the left, each
image shows a successive iteration in the agglomeration procedure. The third row shows the CD
scores for each of these patches, where patch colors in the second row correspond to line colors in
the third row. ACD successfully finds important regions for the target class (such as the puck), and
this importance increases as more pixels are selected. Best viewed in color.",A diagram of various shots of people.,"A detailed layout showing the effects of different types of motion in a race track, including the speed of a bike.","A comprehensive technical explanation of the different types of robotic legs for the human body, including the type of foot and the position of the leg that is placed on the robot in front of the left hand, and the right side.","Figure 3: ACD interpretation for a VGG network prediction, described in 4.2.1. ACD shows that
the CNN is focusing on skates to predict the class “puck”, indicating that the model has captured
dataset bias. The top row shows the original image, logits for the five top-predicted classes, and the
CD superpixel-level scores for those classes. The second row shows separate image patches ACD
has identified as being independently predictive of the class “puck”. Starting from the left, each
image shows a successive iteration in the agglomeration procedure. The third row shows the CD
scores for each of these patches, where patch colors in the second row correspond to line colors in
the third row. ACD successfully finds important regions for the target class (such as the puck), and
this importance increases as more pixels are selected. Best viewed in color.",0.75,0.2671,0.5086,2994,1477,2.027,55e4f7c59f5952911282d45bd3c0a8ab,images/2018/arxiv_0000295.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000296,Figure 296,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure 3: ACD interpretation for a VGG network prediction, described in 4.2.1. ACD shows that
the CNN is focusing on skates to predict the class “puck”, indicating that the model has captured
dataset bias. The top row shows the original image, logits for the five top-predicted classes, and the
CD superpixel-level scores for those classes. The second row shows separate image patches ACD
has identified as being independently predictive of the class “puck”. Starting from the left, each
image shows a successive iteration in the agglomeration procedure. The third row shows the CD
scores for each of these patches, where patch colors in the second row correspond to line colors in
the third row. ACD successfully finds important regions for the target class (such as the puck), and
this importance increases as more pixels are selected. Best viewed in color.",A diagram of a row of toothbrushes.,"A detailed layout showing the structure of a robot in motion with a number of options to choose that shows key features, attributes, and data points.","A comprehensive technical explanation of the use of smart toothbrushes for dental care in the us, 1999 and 2010, with data provided by the u s & p e v company, 2009, and the department of health services.","Figure 3: ACD interpretation for a VGG network prediction, described in 4.2.1. ACD shows that
the CNN is focusing on skates to predict the class “puck”, indicating that the model has captured
dataset bias. The top row shows the original image, logits for the five top-predicted classes, and the
CD superpixel-level scores for those classes. The second row shows separate image patches ACD
has identified as being independently predictive of the class “puck”. Starting from the left, each
image shows a successive iteration in the agglomeration procedure. The third row shows the CD
scores for each of these patches, where patch colors in the second row correspond to line colors in
the third row. ACD successfully finds important regions for the target class (such as the puck), and
this importance increases as more pixels are selected. Best viewed in color.",0.81,0.262,0.536,1224,600,2.04,3b2fb88ee3e8ee9bbf782513453fee1b,images/2018/arxiv_0000296.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000297,Figure 297,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure 4: Results for human studies. A. Binary accuracy for whether a subject correctly selected the
more accurate model using different interpretation techniques B. Average rank (from 1 to 4) of how
much different interpretation techniques helped a subject to trust a model, higher ranks are better.",A diagram of the different levels of.,A detailed layout showing the differences between the two different types of a single - cell and multiple - cell cell.,"A comprehensive technical explanation of the basic results from the two data sets, including the average and most important data sets in the study, as well as shown in figure 1 and 2 / 3 / 4 / 3, and 5.","Figure 4: Results for human studies. A. Binary accuracy for whether a subject correctly selected the
more accurate model using different interpretation techniques B. Average rank (from 1 to 4) of how
much different interpretation techniques helped a subject to trust a model, higher ranks are better.",0.75,0.2184,0.4842,3360,1128,2.979,35d459e50b84dd0cbd9c5f31e46f7b2d,images/2018/arxiv_0000297.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000298,Figure 298,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure 4: Results for human studies. A. Binary accuracy for whether a subject correctly selected the
more accurate model using different interpretation techniques B. Average rank (from 1 to 4) of how
much different interpretation techniques helped a subject to trust a model, higher ranks are better.",A diagram of a bar graph that shows.,"A detailed layout showing the number of different types of computer equipment in the world, including the average and expected.","A comprehensive technical explanation of the average and average passenger time in u s 2010, 2009, 2011, 2013, 2015, 2016, 2017, 2019, 2020, and 202072 - 20, 3895 representing all details, specifications, and configurations of the system components.","Figure 4: Results for human studies. A. Binary accuracy for whether a subject correctly selected the
more accurate model using different interpretation techniques B. Average rank (from 1 to 4) of how
much different interpretation techniques helped a subject to trust a model, higher ranks are better.",0.75,0.2048,0.4774,1224,426,2.873,d45f817b0bad88b96b05b6ae1d287a9d,images/2018/arxiv_0000298.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000299,Figure 299,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure S1: Intuition for CD run on a corner-shaped blob compared to build-up and occlusion. CD
decomposes a DNN’s feedforward pass into a part from the blob of interest (top row) and everything
else (second row). Left column shows original image with overlaid blob. Other columns show DNN
activations summed over the filter dimension. Top and third rows are on same color scale. Second
and bottom rows are on same color scale.",A diagram of a number of different.,"A detailed layout showing the size and structure of the cells in a human cell, with a large number of cells.","A comprehensive technical explanation of the human brain and its functions in a computer system, including the brain and the nervous system, and the brain, from the first to the second to the third to the fourth to the next.","Figure S1: Intuition for CD run on a corner-shaped blob compared to build-up and occlusion. CD
decomposes a DNN’s feedforward pass into a part from the blob of interest (top row) and everything
else (second row). Left column shows original image with overlaid blob. Other columns show DNN
activations summed over the filter dimension. Top and third rows are on same color scale. Second
and bottom rows are on same color scale.",0.7357,0.2436,0.4897,3858,2020,1.91,c6667205eb143724e85d10c50525586c,images/2018/arxiv_0000299.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000300,Figure 300,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure S2: Comparing unit-level CD scores for the correct class to scores from baseline methods.
In each case, the model correctly predicts the label, shown on the y axis. Blue is positive, white is
neutral, and red is negative. Best viewed in color.
Fig S1 gives intuition for CD on the VGG-16 ImageNet model described in Sec 4. CD keeps track
of the contributions of the blob and non-blob throughout the network. This is intuitively similar
to the occlusion and build-up methods, shown in the bottom two rows. The build-up method sets
everything but the patch of interest to a references value (often zero). These rows compare the CD
decomposition to perturbing the input as in the occlusion and build-up methods. They are similar in
early layers, but differences become apparent in later layers. 12",A diagram of different images of birds.,"A detailed layout showing a bird, a lizard and a computer screen with images of it ' s surroundings.","A comprehensive technical explanation of the four different images used to create a bird, a snake and a lizard, and a computer screen with the same image on it ' s screen background color scheme, including a bird ' s head.","Figure S2: Comparing unit-level CD scores for the correct class to scores from baseline methods.
In each case, the model correctly predicts the label, shown on the y axis. Blue is positive, white is
neutral, and red is negative. Best viewed in color.
Fig S1 gives intuition for CD on the VGG-16 ImageNet model described in Sec 4. CD keeps track
of the contributions of the blob and non-blob throughout the network. This is intuitively similar
to the occlusion and build-up methods, shown in the bottom two rows. The build-up method sets
everything but the patch of interest to a references value (often zero). These rows compare the CD
decomposition to perturbing the input as in the occlusion and build-up methods. They are similar in
early layers, but differences become apparent in later layers. 12",0.7071,0.2557,0.4814,562,553,1.016,54d161a29f62728e5531ccf9d9684683,images/2018/arxiv_0000300.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000301,Figure 301,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"SST top-predicted examples.
Here, the model used and figure produced
correspond to Fig 2.",A diagram of a line showing the number.,"A detailed layout showing the number of different types of seats in a stadium seating area, including the stadium seating.","A comprehensive technical explanation of the process for the next project, including the project schedule and the completion process for each project, as well as the final project status and time period of the project description of development and.","SST top-predicted examples.
Here, the model used and figure produced
correspond to Fig 2.",0.75,0.2618,0.5059,3787,746,5.076,5657670ef8259b78d7ded3a98cd69b36,images/2018/arxiv_0000301.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000302,Figure 302,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the number of people in a.,A detailed layout showing the number of employees in each company for each employee to attend and receive a job.,"A comprehensive technical explanation of the current and future market hours for the u s, canada and australia markets, from 2008 to 2015, with the following dates of each month ending in full year, and ending at the same time.",NULL,0.75,0.2356,0.4928,3787,746,5.076,12e3911fc818548c0d0de0372ac7802d,images/2018/arxiv_0000302.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000303,Figure 303,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the number of people in.,A detailed layout showing the number of people who are in each country on a map of numbers and words.,A comprehensive technical explanation of the time and date of the day for an event in which the events are held and where the dates are placed on the table numbers are marked in the rows of the table and the numbers.,NULL,0.7107,0.1875,0.4491,3787,746,5.076,f44586585cd39baebcd35da5601d8fc2,images/2018/arxiv_0000303.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000304,Figure 304,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the time zones for all of.,"A detailed layout showing the number of people in each country on a map, with different numbers and locations.","A comprehensive technical explanation of the different types of water in the world ' s oceans and oceans, including the water cycle and the oceans and the ocean, including oceans and sea, and the temperatures and oceans and seas,.",NULL,0.7429,0.1747,0.4588,3787,746,5.076,18d5c492d1b8d006a98c66bb969bbe9d,images/2018/arxiv_0000304.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000305,Figure 305,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the different types of.,"A detailed layout showing the time of each day and the schedule for each day of the week, with a full schedule.","A comprehensive technical explanation of the operating environment for the new computer system, including the same hardware and peripheral components as well as the current software development capabilities of the system - based system - updating -.",NULL,0.7464,0.2418,0.4941,3787,746,5.076,a44d7a5db5a0b49bdcbf291466b5aa41,images/2018/arxiv_0000305.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000306,Figure 306,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the number of people on a.,A detailed layout showing the numbers of employees and their timeframes for each organization to work.,"A comprehensive technical explanation of the software and its features for the project management process in the u s a r v e o d d e, incl v, inc, is shown here in this diagram, with a 2 representing all details, specifications, and configurations.",NULL,0.7143,0.2294,0.4718,3787,746,5.076,9fc9bb83e9c4910afe78b411340d1620,images/2018/arxiv_0000306.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000307,Figure 307,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"MNIST top-predicted examples.
Here, the model used is the same as in
Sec 4.3.2 and the interpretation of the figure produced is the same as in Fig 3.",A diagram of the different patterns of.,A detailed layout showing the distribution of the number of cells in the cell in which each cell is active.,"A comprehensive technical explanation of the new and older model for the new gene species of the same type of gene to be found in the past, present life - time, and present, in the next generations of the future.","MNIST top-predicted examples.
Here, the model used is the same as in
Sec 4.3.2 and the interpretation of the figure produced is the same as in Fig 3.",0.7921,0.2697,0.5309,3556,1493,2.382,a3c7e134b6dc75ba6612c8cfdc0ae19d,images/2018/arxiv_0000307.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000308,Figure 308,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"SST lowest-predicted examples.
Here, the model used and figure produced
correspond to Fig 2.",A diagram of a table with a number of.,"A detailed layout showing the number of employees at a company ' s office in the united states, including the company '.","A comprehensive technical explanation of the current and future air traffic systems in the united states, from 1953 to 2017 - page 7 of 8 - click to view larger image - page - 1 jp jpg0005 jp jpl representing all details, specifications, and.","SST lowest-predicted examples.
Here, the model used and figure produced
correspond to Fig 2.",0.75,0.1975,0.4738,3787,746,5.076,291c644d2968422b5f6322d2b158bd54,images/2018/arxiv_0000308.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000309,Figure 309,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"SST lowest-predicted examples.
Here, the model used and figure produced
correspond to Fig 2.",A diagram of the number of people on.,A detailed layout showing the numbers of people who are in each state and their numbers on each map.,"A comprehensive technical explanation of the internet security market in a country with a great price range of customers to buy or sell in china for cash and stock options to buy to buy in europe for a second - buy, and buy.","SST lowest-predicted examples.
Here, the model used and figure produced
correspond to Fig 2.",0.7071,0.1867,0.4469,3787,746,5.076,1840d0ddc7f61d534044ea0469bbfba4,images/2018/arxiv_0000309.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000310,Figure 310,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"MNIST top-predicted examples.
Here, the model used is the same as in
Sec 4.3.2 and the interpretation of the figure produced is the same as in Fig 3.",A diagram of the number and time of.,"A detailed layout showing the number of seats in each plane, with the time in each seat to take off.","A comprehensive technical explanation of the cost of a flight in the united states from the airliner ' s perspective to the current airlinerships, and the time of flight times, and location, with the total numbers,.","MNIST top-predicted examples.
Here, the model used is the same as in
Sec 4.3.2 and the interpretation of the figure produced is the same as in Fig 3.",0.7071,0.2401,0.4736,3787,746,5.076,8acf2576fadd2b45365c45cab47654c4,images/2018/arxiv_0000310.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000311,Figure 311,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of different types of data.,A detailed layout showing the differences in the various types of cell phones and the number of cellular phones.,"A comprehensive technical explanation of the data processing technique for the model of the human body, including the distribution of the number of the cells, and the size of the cellphoneus, the cell, and number of its tissue.",NULL,0.75,0.2408,0.4954,3556,1493,2.382,94a5b1ddd59c5c6776d10825315852b9,images/2018/arxiv_0000311.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000312,Figure 312,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the different types of.,"A detailed layout showing the various types of the data for the model, and the different timeframes.","A comprehensive technical explanation of the evolution of the plasmas in the plasma system, including the first two - dimensional model of the system and the next one - dimensional representation of the first - dimensional diagram of the second -.",NULL,0.7671,0.2845,0.5258,3556,1493,2.382,ef858afdecdde8aac3595bf257d6f354,images/2018/arxiv_0000312.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000313,Figure 313,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the different types of.,A detailed layout showing the various types of the data that is generated by the data on the screen.,"A comprehensive technical explanation of the time framer in the data visual library for data visual viewers and data visual tools, including data visual tooling and visual tools and software, v2 / r2 / 2 / 3.",NULL,0.7071,0.2547,0.4809,3563,1493,2.386,5c44c7ead77339f7a6e30c4cea0ea8bf,images/2018/arxiv_0000313.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000314,Figure 314,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"MNIST lowest-predicted examples.
Here, the model used is the same as
in Sec 4.3.2 and the interpretation of the figure produced is the same as in Fig 3.",A diagram of the different types of.,"A detailed layout showing the various stages of a data visual in a computer system, including a number of data.","A comprehensive technical explanation of the different data processing techniques used to produce data for a computer system, including the data processing process and the application of data processing tools in the image outputing process for the.","MNIST lowest-predicted examples.
Here, the model used is the same as
in Sec 4.3.2 and the interpretation of the figure produced is the same as in Fig 3.",0.7464,0.2612,0.5038,3547,1505,2.357,9e4d011aa58313e35418dc9676151fba,images/2018/arxiv_0000314.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000315,Figure 315,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of a line graph that shows.,"A detailed layout showing the number of different objects in the same region, and how they are arranged.","A comprehensive technical explanation of the temperature and humidity of the atmosphere on earth, including the average temperature for each of the three poles of earth ' s orbites, in the atmospheres, in space, and the solars.",NULL,0.7214,0.2436,0.4825,3562,1493,2.386,1f4a3cace224e0b408ca7456f9494fa9,images/2018/arxiv_0000315.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000316,Figure 316,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the different types of.,A detailed layout showing the number of different types of data collected by a single cell phone and also using the same data.,"A comprehensive technical explanation of the various patterns of the human body ' s magnetic field and magnetic field, including the magnets and magnetic fields of the magnet magnets on the magnet and magnets of the magnetic plates,.",NULL,0.75,0.2508,0.5004,3568,1493,2.39,2528ff3876ef52ab56a7c474768e2d19,images/2018/arxiv_0000316.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000317,Figure 317,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Imagenet top-predicted examples.
Here, the model used and figure pro-
duced correspond to that in Fig 3.",A diagram of a number of different.,"A detailed layout showing the various types of the data used in this image, including a bar graph and a line chart.","A comprehensive technical explanation of the different patterns of the current data source with different data sources to be used in the same way, including the data source, the data and the data, on the data visual, the first,.","Imagenet top-predicted examples.
Here, the model used and figure pro-
duced correspond to that in Fig 3.",0.87,0.2539,0.5619,3562,1493,2.386,63911eac6b2c19592a20fc1bfff8856c,images/2018/arxiv_0000317.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000318,Figure 318,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the different kinds of.,A detailed layout showing the various lines of data that are used to create a chart of the different types of data.,"A comprehensive technical explanation of the simulation of the two - dimensional model for the same area, including the same number of objects, and the different areas of the same size of the number of the time of the region and time.",NULL,0.81,0.2483,0.5292,3572,1493,2.392,8d2bc4a3f7bbfb73f4ae8c61cb9f38a4,images/2018/arxiv_0000318.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000319,Figure 319,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of a black and white dog is.,A detailed layout showing the different types of cells in a cell phone and what they are used to make them.,"A comprehensive technical explanation of a polar bear ' s habitat in the wild and what does it mean? - page 1 - page 2 - page 3 of 4 - animal - 1 - png - p2 - p1 representing all details, specifications, and configurations of the system components.",NULL,0.7321,0.2319,0.482,2962,590,5.02,66ae6218bafc0b60bda8a362b3a8bfb7,images/2018/arxiv_0000319.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000320,Figure 320,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of a group of people walking.,A detailed layout showing the distribution of the various groups of people in different groups of color coded area.,"A comprehensive technical explanation of the number of people in a field with different colors and sizes, including numbers of people walking along a line graph, and scatters, and points, and dots, and data, are shown.",NULL,0.81,0.2657,0.5379,2994,1259,2.378,1ad207610761c60de4c0897ea9c20b03,images/2018/arxiv_0000320.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000321,Figure 321,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of a bird with long legs and.,A detailed layout showing the distribution of the size and composition of a bird in different colors and sizes.,"A comprehensive technical explanation of a rabbit ' s diet and its origins in the wild, including the following species and their evolutions and distributions and habitats, with examples of food and habitat changes, from the previous study.",NULL,0.8064,0.2542,0.5303,2989,590,5.066,2c74af3c28b95a91db7b7a687470c20c,images/2018/arxiv_0000321.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000322,Figure 322,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the birds are walking.,A detailed layout showing the distribution of the species of birds in this area and the location of the birds.,"A comprehensive technical explanation of the birds in the field are different colors and sizes of the bird ' s feathers and beaks, with some of which are orange, blue, orange, and one is a blue, red and orange,.",NULL,0.8029,0.2838,0.5433,2994,1249,2.397,5e9e0e4c88c2e6135e6e64b26829119b,images/2018/arxiv_0000322.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000323,Figure 323,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of a bird that is standing.,A detailed layout showing the distribution of human hair and hair growth in the same area of the body.,A comprehensive technical explanation of the human gutter of a zebra in the wild and an image of an ostrich in the grass with a graph of a bar chart to the bottom and a line of the image of the animal.,NULL,0.7743,0.2097,0.492,2962,590,5.02,94e840360ec58fad9f5d9ca0d246d80b,images/2018/arxiv_0000323.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000324,Figure 324,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Imagenet lowest-predicted examples.
Here, the model used and figure
produced correspond to that in Fig ??.",A diagram of a group of people walking.,A detailed layout showing a large group of ostriches in a field with several different colored dots.,"A comprehensive technical explanation of a field plot and the different types of birds in the field for each species of bird species, including ostrich and ostriches and ospreas, or ostrichs, from the ostric.","Imagenet lowest-predicted examples.
Here, the model used and figure
produced correspond to that in Fig ??.",0.7071,0.3105,0.5088,2994,1245,2.405,787d71b961956e559c03a17f5e1bede1,images/2018/arxiv_0000324.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000325,Figure 325,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of an animal with a cell.,"A detailed layout showing the different types of cell phones with their respective colors and size, and the corresponding numbers.","A comprehensive technical explanation of the evolution of the human cell phone in the past and present in the future, from the earliest to the present, with the earliest time, the earliest years of the earliest generations, and the present.",NULL,0.75,0.2011,0.4756,2962,590,5.02,0752ebc2d0e975f0cd68f35b908ff002,images/2018/arxiv_0000325.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000326,Figure 326,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the number of different.,"A detailed layout showing the number and type of items used for the project, including the numbers of each item.","A comprehensive technical explanation of the various types of the data that are used to determine the exact time for a project or project, including the data and the results of the following the results invoice of the results,.",NULL,0.75,0.2483,0.4991,3002,1238,2.425,5bd0d6c4012b80a2eff09ec5433633ce,images/2018/arxiv_0000326.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000327,Figure 327,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the same image of a boat.,"A detailed layout showing the number of reeds in different stages of growth, and the size of each species.","A comprehensive technical explanation of the water cycle in the natural world, including one of which is a lake and two of which are in a row with a boat and a fish in a log and a row on it representing all details, specifications, and.",NULL,0.7286,0.2365,0.4826,2977,590,5.046,6c0742e94a7000bd1e3e5afcf4f2cd21,images/2018/arxiv_0000327.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000328,Figure 328,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the cycle path for people.,A detailed layout showing the number of bikes in different stages of development and the number by year of use.,"A comprehensive technical explanation of the cycle cycle of the cyclist in different stages of development and development, with data lines showing the cycle cycles in each phase, and the cycle, as well as well, and at the same time.",NULL,0.7464,0.3173,0.5318,3042,1241,2.451,6cb47a01014f77c50f96d7a3b6471d4e,images/2018/arxiv_0000328.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000329,Figure 329,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the structure of an.,A detailed layout showing the effects of light pollution and radiation in the atmosphere of earth ' s atmosphere.,"A comprehensive technical explanation of the structure of a large cell phone tower, including a long, thin wall with a number of columns and a vertical line of smaller lines on it ' s that are numbered ' s ' s.",NULL,0.75,0.2199,0.4849,2962,590,5.02,fa0d1a83b88783bee880424f4a1ab5a5,images/2018/arxiv_0000329.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000330,Figure 330,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of the various color.,A detailed layout showing the number of different lines in a data visual chart with a black background and a blue line.,"A comprehensive technical explanation of the data visual for the gis - based data visual system, including data visual and data visual elements, in the data gis and data modeling process, in one image, and in the same.",NULL,0.81,0.2742,0.5421,3042,1256,2.422,eddc352ed065411b705205a0fb5e623f,images/2018/arxiv_0000330.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000331,Figure 331,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure S3: Example of ACD run on an image of class 0 before and after an adversarial perturbation
(a DeepFool attack). Best viewed in color.",A diagram of the different types of.,A detailed layout showing the different temperatures and temperatures of snow in various regions of the earth.,"A comprehensive technical explanation of the differences between the two data visual systems in the same area of the system, including the data visual cluster and the data flow from the data source to the data consoles on the data system.","Figure S3: Example of ACD run on an image of class 0 before and after an adversarial perturbation
(a DeepFool attack). Best viewed in color.",0.7429,0.2635,0.5032,3568,1493,2.39,9b8eddac38255dba343f7d3ac4d16546,images/2018/arxiv_0000331.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000332,Figure 332,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,NULL,A diagram of a line graph shows the.,"A detailed layout showing the time of a person to get out of bed, and a line graph of time that shows key features, attributes, and data points in.","A comprehensive technical explanation of the different types of data processing systems in a computer system, including the data processing process and the modeling process for the model of datas to be used in the model for the machine - outflow.",NULL,0.87,0.2505,0.5603,3556,1493,2.382,8a5655dfcc4ce86d6d21d15bf46e7a51,images/2018/arxiv_0000332.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000333,Figure 333,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure S4: Examples of attacks for one image. Original image (left column) is correctly predicted as
class 0. After each adversarial perturbation (middle column), the predicted class for the adversarial
image (right column) is now altered. S7
GENERALIZING CD TO CNNS",A diagram of the different image and.,A detailed layout showing the number of images that are being displayed on the page in the text box.,"A comprehensive technical explanation of the original image versusarinal perturatativeal image in a single cell, with an unconcented image of a single source at some possible color and multiple color to one, two - dimensional.","Figure S4: Examples of attacks for one image. Original image (left column) is correctly predicted as
class 0. After each adversarial perturbation (middle column), the predicted class for the adversarial
image (right column) is now altered. S7
GENERALIZING CD TO CNNS",0.7071,0.2213,0.4642,642,1101,0.583,4ace577cb8a2e93fbd3db35b4e824d98,images/2018/arxiv_0000333.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000334,Figure 334,scientific_figure,Hierarchical interpretations for neural network predictions,arXiv Research Authors,1806.05337v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Hierarchical interpretations for neural network predictions. arXiv:1806.05337v2,"Figure S5: Comparing unit-level CD scores to CD scores from the naive extension of CD to CNNs,
independently developed by Godin et al. (2018). Labels under the bottom row signify the minimum
and maximum scores from each column. Altering the bias partition and ReLU decomposition qual-
itatively improves scores (e.g. see scores in bottom row corresponding to the location of the crane),
and avoids extremely large magnitudes (see values under left two columns). Blue is positive, white
is neutral, and red is negative. In each case, scores are for the correct class, which the model predicts
correctly (shown on the y axis).",A diagram of the different stages of a.,"A detailed layout showing the differences in the size of a bird ' s wing and the amount of wings that shows key features, attributes, and data.","A comprehensive technical explanation of the different levels of heat in the environment and what is it like to use it?, and how does it work???,?? and how to do you?? - image comparison representing all details, specifications, and configurations.","Figure S5: Comparing unit-level CD scores to CD scores from the naive extension of CD to CNNs,
independently developed by Godin et al. (2018). Labels under the bottom row signify the minimum
and maximum scores from each column. Altering the bias partition and ReLU decomposition qual-
itatively improves scores (e.g. see scores in bottom row corresponding to the location of the crane),
and avoids extremely large magnitudes (see values under left two columns). Blue is positive, white
is neutral, and red is negative. In each case, scores are for the correct class, which the model predicts
correctly (shown on the y axis).",0.81,0.1944,0.5022,766,568,1.349,8ee7e2063af8f1d83f2d6d47fabdc436,images/2018/arxiv_0000334.png,https://arxiv.org/pdf/1806.05337v2.pdf
arxiv_0000335,Figure 335,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,"Figure 1: Traditional transfer learning versus our new transfer learning framework. Instead of transferring
features, we transfer the graphs output by a network. The graphs are multiplied by task-specific features (e.g.
embeddings or hidden states) to produce structure-aware features.",A diagram of a table and two tables.,A detailed layout showing the basic components of a typical timeframes and its functions in a single - dimensional.,"A comprehensive technical explanation of the application of the data processing system for a single - dimensional data transfer system, from a to b and c, with a, e, and b, t, and, i, t t representing all details, specifications, and configurations.","Figure 1: Traditional transfer learning versus our new transfer learning framework. Instead of transferring
features, we transfer the graphs output by a network. The graphs are multiplied by task-specific features (e.g.
embeddings or hidden states) to produce structure-aware features.",0.75,0.301,0.5255,832,299,2.783,d1fc326e6c17af31f1d12afd59dd2ac1,images/2018/arxiv_0000335.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000336,Figure 336,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,"Figure 2: Overview of our approach GLoMo. During the unsupervised learning phase, the feature predictor
and the graph predictor are jointly trained to perform context prediction. During the transfer phase, the graph
predictor is frozen and used to extract graphs for the downstream tasks. An RNN decoder is applied to all
positions in the feature predictor, but we only show the one at position “A” for simplicity. “Select one” means
the graphs can be transferred to any layer in the downstream task model. “FF” refers to feed-forward networks.
The graphs output by the graph predictor are used as the weights in the “weighted sum” operation (see Eq. 2).",A diagram of an open - source data.,"A detailed layout showing the multiple paths of an open - source dns system, including a single - line.","A comprehensive technical explanation of the new and expanded wan architecture, with its components and architectures, and their examples, for each of the following versions, as well, as far as well as well on the following, as.","Figure 2: Overview of our approach GLoMo. During the unsupervised learning phase, the feature predictor
and the graph predictor are jointly trained to perform context prediction. During the transfer phase, the graph
predictor is frozen and used to extract graphs for the downstream tasks. An RNN decoder is applied to all
positions in the feature predictor, but we only show the one at position “A” for simplicity. “Select one” means
the graphs can be transferred to any layer in the downstream task model. “FF” refers to feed-forward networks.
The graphs output by the graph predictor are used as the weights in the “weighted sum” operation (see Eq. 2).",0.7179,0.3064,0.5121,833,367,2.27,2c6b3e1575b1155fa7cae63c915e1b4d,images/2018/arxiv_0000336.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000337,Figure 337,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,"Figure 3: Visualization of the graphs on the MNLI dataset. The graph predictor has not been trained on MNLI.
The words on the y-axis “attend” to the words on the a-axis; i.e., each row sums to 1.",A diagram of the number of people who.,A detailed layout showing the number of families and the percentage of children who have been born to a family.,"A comprehensive technical explanation of the average number of people who have visited the region in their family history and age groups, by region and region, 2010 - 2011 to 2016 - 2017 - 2020 - 2019 - 19 - 19 png.","Figure 3: Visualization of the graphs on the MNLI dataset. The graph predictor has not been trained on MNLI.
The words on the y-axis “attend” to the words on the a-axis; i.e., each row sums to 1.",0.7464,0.2445,0.4954,1116,1116,1.0,4d7473fd4a208c77bcc903278f59cb74,images/2018/arxiv_0000337.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000338,Figure 338,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,Attending to objects for modeling long-term dependency.,A diagram of the number of people who.,"A detailed layout showing the distribution of the number of people who live in the region, by population.","A comprehensive technical explanation of the study of the number of individuals in greengrowns and undergrowthing the vegetation, including where they are related to be related in glassy - related to the glassy, and the water.",Attending to objects for modeling long-term dependency.,0.785,0.2347,0.5099,1137,1137,1.0,3b24428f94b98be662ab1ba8868fe521,images/2018/arxiv_0000338.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000339,Figure 339,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,Attending to negative words and predicates.,A diagram of a bar chart with the.,A detailed layout showing the number of pages that are written in a bar chart with the same numbers.,"A comprehensive technical explanation of the israeli press and its historical history, especially in - - - between - age and age, in the past, and present, in - present, with their own - present - present and present.",Attending to negative words and predicates.,0.7671,0.274,0.5206,1108,1108,1.0,52d9e16b27f3fb81e46af979f04bb7f4,images/2018/arxiv_0000339.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000340,Figure 340,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,NULL,A diagram of the number of people in.,A detailed layout showing the number of people who have been selected in the survey by the department of the environment.,"A comprehensive technical explanation of the data visual for the internet web users and their users, based on google search engine results and the results of the results on the web page below from the linkpages and the webs p.",NULL,0.75,0.2445,0.4972,1119,1119,1.0,44dd6cf2eac9c17141100ef1805981d0,images/2018/arxiv_0000340.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000341,Figure 341,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,"Figure 3: Visualization of the graphs on the MNLI dataset. The graph predictor has not been trained on MNLI.
The words on the y-axis “attend” to the words on the a-axis; i.e., each row sums to 1.",A diagram of the different stages of a.,A detailed layout showing the distribution of the human body for a variety of different types of cell lines.,"A comprehensive technical explanation of the role of the human cell in the body ' s first - quarter of the year, and the results of the second - year - end - end of - stage - year, a - year representing all details, specifications, and.","Figure 3: Visualization of the graphs on the MNLI dataset. The graph predictor has not been trained on MNLI.
The words on the y-axis “attend” to the words on the a-axis; i.e., each row sums to 1.",0.7957,0.2432,0.5194,1224,978,1.252,753ff1499af87d20b0ebec37c140c31c,images/2018/arxiv_0000341.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000342,Figure 342,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,"Figure 4: Visualization. Left: a shark image as the input. Middle: weights of the edges connected
with the central pixel, organized into 24 heads (3 layers with 8 heads each). Right: weights of the
edges connected with the bottom-right pixel. Note the use of masking. Method / Base-model
ResNet-18
ResNet-34 baseline
90.93±0.33
91.42±0.17 GLoMo
91.55±0.23
91.70±0.09 ablation: uniform graph
91.07±0.24
-",A diagram of the various types of hair.,A detailed layout showing the different stages of the hand drawn lines and shapes in the form of a bird.,"A comprehensive technical explanation of the initial of an airplane flight and its flight path, including the landing gear and landing gear, and the engine and engine compartment of the engine, which is the engine are located on the main body.","Figure 4: Visualization. Left: a shark image as the input. Middle: weights of the edges connected
with the central pixel, organized into 24 heads (3 layers with 8 heads each). Right: weights of the
edges connected with the bottom-right pixel. Note the use of masking. Method / Base-model
ResNet-18
ResNet-34 baseline
90.93±0.33
91.42±0.17 GLoMo
91.55±0.23
91.70±0.09 ablation: uniform graph
91.07±0.24
-",0.7214,0.2946,0.508,699,290,2.41,44aa7ca319bc16953204b41287d426c3,images/2018/arxiv_0000342.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000343,Figure 343,scientific_figure,GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations,arXiv Research Authors,1806.05662v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). GLoMo: Unsupervisedly Learned Relational Graphs as Transferable Representations. arXiv:1806.05662v3,"Figure 4: Visualization. Left: a shark image as the input. Middle: weights of the edges connected
with the central pixel, organized into 24 heads (3 layers with 8 heads each). Right: weights of the
edges connected with the bottom-right pixel. Note the use of masking.",A diagram of a large group of birds.,A detailed layout showing a group of birds flying over a blue background with white clouds and a single plane.,"A comprehensive technical explanation of the use of the sound wave in music, including notes and pictures of the ocean and the sky, and how it affects to understand what sounds are used for the sound waves and how they can be.","Figure 4: Visualization. Left: a shark image as the input. Middle: weights of the edges connected
with the central pixel, organized into 24 heads (3 layers with 8 heads each). Right: weights of the
edges connected with the bottom-right pixel. Note the use of masking.",0.7429,0.2927,0.5178,1224,300,4.08,22164b91fea0e722476ee4041c05a858,images/2018/arxiv_0000343.png,https://arxiv.org/pdf/1806.05662v3.pdf
arxiv_0000344,Figure 344,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 1: Illustration of the proposed Multimodal Factorization Model (MFM) with three modalities. MFM
factorizes multimodal representations into multimodal discriminative factors Fy and modality-specific generative
factors Fa{1∶M}. (a) MFM Generative Network with latent variables {Zy, Za{1∶M}}, factors {Fy, Fa{1∶M}},
generated multimodal data ˆX1∶3 and labels ˆY. (b) MFM Inference Network. (c) MFM Neural Architecture.
Best viewed zoomed in and in color. discriminative information (Srivastava & Salakhutdinov, 2012). To sum up, MFM defines a joint
distribution over multimodal data, and by the conditional independence assumptions in the assumed
graphical model, both generative and discriminative aspects are taken into account. Our model design
further provides interpretability of the factorized representations. Through an extensive set of experiments, we show that MFM learns improved multimodal representa-
tions with these characteristics: 1) The multimodal discriminative factors achieve state-of-the-art or
competitive performance on six multimodal time series datasets. We also demonstrate that MFM can
generalize by integrating it with other existing multimodal discriminative models. 2) MFM allows
flexible generation concerning multimodal discriminative factors (labels) and modality-specific gener-
ative factors (styles). We further show that we can perform reconstruction of missing modalities from
observed modalities without significantly impacting discriminative performance. Finally, we interpret
our learned representations using information-based and gradient-based methods, allowing us to
understand the contributions of individual factors towards multimodal prediction and generation. 2
MULTIMODAL FACTORIZATION MODEL Multimodal Factorization Model (MFM) is a latent variable model (Figure 1(a)) with conditional
independence assumptions over multimodal discriminative factors and modality-specific generative
factors. According to these assumptions, we propose a factorization over the joint distribution of
multimodal data (Section 2.1). Since exact posterior inference on this factorized distribution can be
intractable, we propose an approximate inference algorithm based on minimizing a joint-distribution
Wasserstein distance over multimodal data (Section 2.2). Finally, we derive the MFM objective by
approximating the joint-distribution Wasserstein distance via a generalized mean-field assumption. Notation: We define X1∶M as the multimodal data from M modalities and Y as the labels, with joint
distribution PX1∶M,Y = P(X1∶M,Y). Let ˆX1∶M denote the generated multimodal data and ˆY denote
the generated labels, with joint distribution P ˆ
X1∶M, ˆ
Y = P(ˆX1∶M, ˆY). 2.1
FACTORIZED MULTIMODAL REPRESENTATIONS To factorize multimodal representations into multimodal discriminative factors and modality-specific
generative factors, MFM assumes a Bayesian network structure as shown in Figure 1(a). In this
graphical model, factors Fy and Fa{1∶M} are generated from mutually independent latent variables
Z = [Zy,Za{1∶M}] with prior PZ. In particular, Zy generates the multimodal discriminative factor Fy
and Za{1∶M} generate modality-specific generative factors Fa{1∶M}. By construction, Fy contributes
to the generation of ˆY while {Fy,Fai} both contribute to the generation of ˆXi. As a result, the joint
distribution P(ˆX1∶M, ˆY) can be factorized as follows: P ( ˆ
X1∶M, ˆ
Y) = ∫F,Z P ( ˆ
X1∶M, ˆ
Y∣F)P (F∣Z)P (Z)dFdZ (P ( ˆ
Y∣Fy)
M
∏
i=1
P ( ˆ
Xi∣Fai, Fy))(P (Fy∣Zy)
M
∏
i=1
P (Fai∣Zai))(P (Zy)
M
∏
i=1
P (Zai))dFdZ,
(1) = ∫Fy,Fa{1∶M}
Zy,Za{1∶M} with dF = dFy ∏M
i=1 dFai and dZ = dZy ∏M
i=1 dZai. Exact posterior inference in Equation 1 may be analytically intractable due to the integration over Z.
We therefore resort to using an approximate inference distribution Q(Z∣X1∶M,Y) as detailed in the 2",A diagram of a computer keyboard with.,"A detailed layout showing the different functions of a computer keyboard and mouse keyboard, with the keys highlighted.","A comprehensive technical explanation of the multimode fusione and encodes for the first time in a series of three years, each with a different functions to be different size and length and number of the same size,.","Figure 1: Illustration of the proposed Multimodal Factorization Model (MFM) with three modalities. MFM
factorizes multimodal representations into multimodal discriminative factors Fy and modality-specific generative
factors Fa{1∶M}. (a) MFM Generative Network with latent variables {Zy, Za{1∶M}}, factors {Fy, Fa{1∶M}},
generated multimodal data ˆX1∶3 and labels ˆY. (b) MFM Inference Network. (c) MFM Neural Architecture.
Best viewed zoomed in and in color. discriminative information (Srivastava & Salakhutdinov, 2012). To sum up, MFM defines a joint
distribution over multimodal data, and by the conditional independence assumptions in the assumed
graphical model, both generative and discriminative aspects are taken into account. Our model design
further provides interpretability of the factorized representations. Through an extensive set of experiments, we show that MFM learns improved multimodal representa-
tions with these characteristics: 1) The multimodal discriminative factors achieve state-of-the-art or
competitive performance on six multimodal time series datasets. We also demonstrate that MFM can
generalize by integrating it with other existing multimodal discriminative models. 2) MFM allows
flexible generation concerning multimodal discriminative factors (labels) and modality-specific gener-
ative factors (styles). We further show that we can perform reconstruction of missing modalities from
observed modalities without significantly impacting discriminative performance. Finally, we interpret
our learned representations using information-based and gradient-based methods, allowing us to
understand the contributions of individual factors towards multimodal prediction and generation. 2
MULTIMODAL FACTORIZATION MODEL Multimodal Factorization Model (MFM) is a latent variable model (Figure 1(a)) with conditional
independence assumptions over multimodal discriminative factors and modality-specific generative
factors. According to these assumptions, we propose a factorization over the joint distribution of
multimodal data (Section 2.1). Since exact posterior inference on this factorized distribution can be
intractable, we propose an approximate inference algorithm based on minimizing a joint-distribution
Wasserstein distance over multimodal data (Section 2.2). Finally, we derive the MFM objective by
approximating the joint-distribution Wasserstein distance via a generalized mean-field assumption. Notation: We define X1∶M as the multimodal data from M modalities and Y as the labels, with joint
distribution PX1∶M,Y = P(X1∶M,Y). Let ˆX1∶M denote the generated multimodal data and ˆY denote
the generated labels, with joint distribution P ˆ
X1∶M, ˆ
Y = P(ˆX1∶M, ˆY). 2.1
FACTORIZED MULTIMODAL REPRESENTATIONS To factorize multimodal representations into multimodal discriminative factors and modality-specific
generative factors, MFM assumes a Bayesian network structure as shown in Figure 1(a). In this
graphical model, factors Fy and Fa{1∶M} are generated from mutually independent latent variables
Z = [Zy,Za{1∶M}] with prior PZ. In particular, Zy generates the multimodal discriminative factor Fy
and Za{1∶M} generate modality-specific generative factors Fa{1∶M}. By construction, Fy contributes
to the generation of ˆY while {Fy,Fai} both contribute to the generation of ˆXi. As a result, the joint
distribution P(ˆX1∶M, ˆY) can be factorized as follows: P ( ˆ
X1∶M, ˆ
Y) = ∫F,Z P ( ˆ
X1∶M, ˆ
Y∣F)P (F∣Z)P (Z)dFdZ (P ( ˆ
Y∣Fy)
M
∏
i=1
P ( ˆ
Xi∣Fai, Fy))(P (Fy∣Zy)
M
∏
i=1
P (Fai∣Zai))(P (Zy)
M
∏
i=1
P (Zai))dFdZ,
(1) = ∫Fy,Fa{1∶M}
Zy,Za{1∶M} with dF = dFy ∏M
i=1 dFai and dZ = dZy ∏M
i=1 dZai. Exact posterior inference in Equation 1 may be analytically intractable due to the integration over Z.
We therefore resort to using an approximate inference distribution Q(Z∣X1∶M,Y) as detailed in the 2",0.75,0.2537,0.5019,3103,1188,2.612,503cc3e63912f9584cbec96333975a84,images/2018/arxiv_0000344.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000345,Figure 345,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,MFM Inference Network.,A diagram of the letters are all white.,A detailed layout showing the letters of different languages on a keyboard keyboard keyboard keys are highlighted in blue.,"A comprehensive technical explanation of the alphabets used by computer engineers to learn how to code for computers and their computers, including the computer keyboard and the keyboard keys and numbers on the keyboard, as well as shown in the.",MFM Inference Network.,0.75,0.2792,0.5146,1921,1016,1.891,17c9e08199c1cd264325767be07d3f58,images/2018/arxiv_0000345.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000346,Figure 346,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 1: Illustration of the proposed Multimodal Factorization Model (MFM) with three modalities. MFM
factorizes multimodal representations into multimodal discriminative factors Fy and modality-specific generative
factors Fa{1∶M}. (a) MFM Generative Network with latent variables {Zy, Za{1∶M}}, factors {Fy, Fa{1∶M}},
generated multimodal data ˆX1∶3 and labels ˆY. (b) MFM Inference Network. (c) MFM Neural Architecture.
Best viewed zoomed in and in color.",A diagram of the different types of.,"A detailed layout showing the various architectures of a network and its connections to different devices, including networked.","A comprehensive technical explanation of the network architecture of an internet network, including the internet network and the internet networking network, as well as the internet web browsers and the webpagednetwork network network network, from.","Figure 1: Illustration of the proposed Multimodal Factorization Model (MFM) with three modalities. MFM
factorizes multimodal representations into multimodal discriminative factors Fy and modality-specific generative
factors Fa{1∶M}. (a) MFM Generative Network with latent variables {Zy, Za{1∶M}}, factors {Fy, Fa{1∶M}},
generated multimodal data ˆX1∶3 and labels ˆY. (b) MFM Inference Network. (c) MFM Neural Architecture.
Best viewed zoomed in and in color.",0.87,0.2819,0.5759,754,231,3.264,551622508fe37d28cbd3197c2874fe7f,images/2018/arxiv_0000346.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000347,Figure 347,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Method
UM(SVHN)
UM(MNIST)
MM
MFM",A diagram of the four different neural.,A detailed layout showing the different types of neural networks that are connected to each other and connected by two separate.,"A comprehensive technical explanation of the neural network and its functions in the computer system, from wikipediart to wifi, including fa2, fa7, fy, f, f2, f5, f4, and f3 representing all details, specifications, and configurations of the system.","Method
UM(SVHN)
UM(MNIST)
MM
MFM",0.81,0.3684,0.5892,1327,707,1.877,f0b6d51e38d8a629b0172ba7abc4ea82,images/2018/arxiv_0000347.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000348,Figure 348,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,Fix 𝐙𝐲,A diagram of numbers and symbols are.,A detailed layout showing the numbers and symbols of different sizes and sizes of tiles with a grid pattern.,"A comprehensive technical explanation of the numbers in the sky and earth ' s surface are made up of pixels, and the image is a grid of numbers on the same background, with a blue and a grid, and a brown.",Fix 𝐙𝐲,0.7357,0.3416,0.5387,342,342,1.0,f82fff74af0837e0606da2894b3e8fff,images/2018/arxiv_0000348.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000349,Figure 349,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,Fix 𝐙𝐚#,A diagram of numbers on a black.,A detailed layout showing the numbers of digits arranged in a random pattern with a black background and white letters.,"A comprehensive technical explanation of the number system in the computer science library, including the scientific elements and the calculations of the numbers and digits on the screen below it are also the symbols, which are not all the numbers.",Fix 𝐙𝐚#,0.75,0.3501,0.5501,527,528,0.998,506db8c05a754a999824d8ca291b39bf,images/2018/arxiv_0000349.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000350,Figure 350,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 3: Models used in the ablation studies of MFM. Each model removes a design component from our
model. Modality reconstruction and sentiment prediction results are reported on CMU-MOSI with best results
in bold. Factorizing multimodal representations into multimodal discriminative factors and modality-specific
generative factors are crucial for improved performance.",A diagram of the different types of.,"A detailed layout showing the number of possible possible locations for different mechanisms in the system, including the time.","A comprehensive technical explanation of the mechanism of an inferctional protein - based protein - bound protein - packed proteined protein - protein protein - loaded protein system c4uper, c3, m4h1.","Figure 3: Models used in the ablation studies of MFM. Each model removes a design component from our
model. Modality reconstruction and sentiment prediction results are reported on CMU-MOSI with best results
in bold. Factorizing multimodal representations into multimodal discriminative factors and modality-specific
generative factors are crucial for improved performance.",0.75,0.2865,0.5182,832,199,4.181,cc31a10d5496a04a090455bd49624f3c,images/2018/arxiv_0000350.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000351,Figure 351,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 4: Analyzing the multimodal represen-
tations learnt in MFM via information-based
(entire dataset) and gradient-based interpreta-
tion methods (single video) on CMU-MOSI.",A diagram of a line graph with a red.,"A detailed layout showing the different lines that appear to be drawn in blue, red and green, and the same line with the same colored line.","A comprehensive technical explanation of the line graph for the first half of the year, with a green arrow on the right side and red line on the left side of the blue line on a white background, the right, indicating the line.","Figure 4: Analyzing the multimodal represen-
tations learnt in MFM via information-based
(entire dataset) and gradient-based interpreta-
tion methods (single video) on CMU-MOSI.",0.75,0.3037,0.5269,800,348,2.299,8b8836c030910b47bf3bd5dcf4928f6d,images/2018/arxiv_0000351.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000352,Figure 352,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 4: Analyzing the multimodal represen-
tations learnt in MFM via information-based
(entire dataset) and gradient-based interpreta-
tion methods (single video) on CMU-MOSI.",A diagram of a table with numbers and.,"A detailed layout showing the table with the date of publication as a conference paper at icr 2019 that shows key features, attributes, and data.","A comprehensive technical explanation of icr 2010 - page 3 of 5, copyright, cenga com, inc, icr, 2009, icrc, icri, icrt, icra, icrs, icrr representing all details, specifications, and configurations of the system components in full context with.","Figure 4: Analyzing the multimodal represen-
tations learnt in MFM via information-based
(entire dataset) and gradient-based interpreta-
tion methods (single video) on CMU-MOSI.",0.87,0.3555,0.6128,417,213,1.958,917f548b7e8536c7c89af8d0cb1bc247,images/2018/arxiv_0000352.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000353,Figure 353,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"G
SURROGATE INFERENCE GRAPHICAL MODEL",A diagram of a multimodal fusion by.,"A detailed layout showing the multiple stages of a multimoal fusion in a single cell phone system that shows key features, attributes, and data.","A comprehensive technical explanation of the multimodal fusion between two different types of cellular signaling systems, including the two - way mechanisms to create a multimodeligns for the same cell structure of a multiplemodal.","G
SURROGATE INFERENCE GRAPHICAL MODEL",0.81,0.2951,0.5525,800,640,1.25,55ef4df95ba59cdd270cbce3952a09bf,images/2018/arxiv_0000353.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000354,Figure 354,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 5: Recurrent neural architecture for MFM. The encoder Q(Zy∣X1∶M) can be parametrized by
any model that performs multimodal fusion (Nojavanasghari et al., 2016; Zadeh et al., 2018a). We
use encoder LSTM networks and decoder LSTM networks (Cho et al., 2014) to parametrize functions
Q(Za1∶M∣X1∶M) and F1∶M respectively, and FCNNs to parametrize functions Gy, Ga{1∶M} and D.",A diagram of the structure of a.,A detailed layout showing the different processes of the process in which the user is able to create and execute.,"A comprehensive technical explanation of the structure and design of an electronic system for micro - devices, including the new generation of embedded devices, and the new development of the first generation of the computer systems - page - 2 -.","Figure 5: Recurrent neural architecture for MFM. The encoder Q(Zy∣X1∶M) can be parametrized by
any model that performs multimodal fusion (Nojavanasghari et al., 2016; Zadeh et al., 2018a). We
use encoder LSTM networks and decoder LSTM networks (Cho et al., 2014) to parametrize functions
Q(Za1∶M∣X1∶M) and F1∶M respectively, and FCNNs to parametrize functions Gy, Ga{1∶M} and D.",0.75,0.2566,0.5033,1224,1195,1.024,db7999254add7754531e4bdceb503e36,images/2018/arxiv_0000354.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000355,Figure 355,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 6: The surrogate inference graphical model to deal with missing modalities in MFM. Red
lines denote original inference in MFM and green lines denote surrogate inference to infer latent
codes given present modalities.",A diagram of a group of four circles.,"A detailed layout showing the four steps of the x - z and z - z system with arrows that shows key features, attributes, and data points in detail.","A comprehensive technical explanation of the z - 3 and z - 4 algorithms for the game of chess, by david l smith, mdi and j m d, incl, ltd, 2009, inc, d representing all details, specifications, and configurations of the system components in full.","Figure 6: The surrogate inference graphical model to deal with missing modalities in MFM. Red
lines denote original inference in MFM and green lines denote surrogate inference to infer latent
codes given present modalities.",0.81,0.2964,0.5532,398,341,1.167,f5f2e99ae4727e6c0aae576a70b75d24,images/2018/arxiv_0000355.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000356,Figure 356,scientific_figure,Learning Factorized Multimodal Representations,arXiv Research Authors,1806.06176v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Factorized Multimodal Representations. arXiv:1806.06176v3,"Figure 6: The surrogate inference graphical model to deal with missing modalities in MFM. Red
lines denote original inference in MFM and green lines denote surrogate inference to infer latent
codes given present modalities.",A diagram of the process of making a.,"A detailed layout showing the different components and functions of an automated device for a computer system, including two computers.","A comprehensive technical explanation of the proposed system for the solar system, including a solar - powered system and an energiezeration of the earth ' s atmosphere ' s orbit ' s sun ' s surface ' s space.","Figure 6: The surrogate inference graphical model to deal with missing modalities in MFM. Red
lines denote original inference in MFM and green lines denote surrogate inference to infer latent
codes given present modalities.",0.75,0.1822,0.4661,1224,316,3.873,9740b7b08ca4625fd824328830c3da9f,images/2018/arxiv_0000356.png,https://arxiv.org/pdf/1806.06176v3.pdf
arxiv_0000357,Figure 357,scientific_figure,DARTS: Differentiable Architecture Search,arXiv Research Authors,1806.09055v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). DARTS: Differentiable Architecture Search. arXiv:1806.09055v2,"Figure 1: An overview of DARTS: (a) Operations on the edges are initially unknown. (b) Continuous
relaxation of the search space by placing a mixture of candidate operations on each edge. (c) Joint
optimization of the mixing probabilities and the network weights by solving a bilevel optimization
problem. (d) Inducing the final architecture from the learned mixing probabilities.",A diagram of the different types of.,A detailed layout showing the various connections of the current and the current wires in each circuitry unit.,"A comprehensive technical explanation of electrical wiring for a home, including the electrical and electrical devices, and the electrical components, and their applications, etc, rh, e g, r, b, c, e, b.","Figure 1: An overview of DARTS: (a) Operations on the edges are initially unknown. (b) Continuous
relaxation of the search space by placing a mixture of candidate operations on each edge. (c) Joint
optimization of the mixing probabilities and the network weights by solving a bilevel optimization
problem. (d) Inducing the final architecture from the learned mixing probabilities.",0.7429,0.2587,0.5008,1244,565,2.202,9564005b4038ce0c273b7e968339e0d8,images/2018/arxiv_0000357.png,https://arxiv.org/pdf/1806.09055v2.pdf
arxiv_0000358,Figure 358,scientific_figure,DARTS: Differentiable Architecture Search,arXiv Research Authors,1806.09055v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). DARTS: Differentiable Architecture Search. arXiv:1806.09055v2,"Figure 1: An overview of DARTS: (a) Operations on the edges are initially unknown. (b) Continuous
relaxation of the search space by placing a mixture of candidate operations on each edge. (c) Joint
optimization of the mixing probabilities and the network weights by solving a bilevel optimization
problem. (d) Inducing the final architecture from the learned mixing probabilities.",A diagram of the various electrical.,"A detailed layout showing the different types of wires used in the electronics project, including the cable,.","A comprehensive technical explanation of parallel and parallel switch types for leds and lamps, including the different types of switches and sockets, and their uses and functions, for each type of a particular purpose, in the same device,.","Figure 1: An overview of DARTS: (a) Operations on the edges are initially unknown. (b) Continuous
relaxation of the search space by placing a mixture of candidate operations on each edge. (c) Joint
optimization of the mixing probabilities and the network weights by solving a bilevel optimization
problem. (d) Inducing the final architecture from the learned mixing probabilities.",0.7393,0.2857,0.5125,1224,457,2.678,94381d569bd55899948af9cda211b217,images/2018/arxiv_0000358.png,https://arxiv.org/pdf/1806.09055v2.pdf
arxiv_0000359,Figure 359,scientific_figure,DARTS: Differentiable Architecture Search,arXiv Research Authors,1806.09055v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). DARTS: Differentiable Architecture Search. arXiv:1806.09055v2,"Figure 3: Search progress of DARTS for convolutional cells on CIFAR-10 and recurrent cells on
Penn Treebank. We keep track of the most recent architectures over time. Each architecture snapshot
is re-trained from scratch using the training set (for 100 epochs on CIFAR-10 and for 300 epochs on
PTB) and then evaluated on the validation set. For each task, we repeat the experiments for 4 times
with different random seeds, and report the median and the best (per run) validation performance of
the architectures over time. As references, we also report the results (under the same evaluation setup;
with comparable number of parameters) of the best existing cells discovered using RL or evolution,
including NASNet-A (Zoph et al., 2018) (2000 GPU days), AmoebaNet-A (3150 GPU days) (Real
et al., 2018) and ENAS (0.5 GPU day) (Pham et al., 2018b).",A diagram of the various phases of the.,A detailed layout showing the different results of each type of protein in a proteined proteined animal.,"A comprehensive technical explanation of the gart - trm and gartt - trp - tra - trc - trs - trx data and trac - trac t - trasm tra representing all details, specifications, and configurations of the system components in full context with annotated.","Figure 3: Search progress of DARTS for convolutional cells on CIFAR-10 and recurrent cells on
Penn Treebank. We keep track of the most recent architectures over time. Each architecture snapshot
is re-trained from scratch using the training set (for 100 epochs on CIFAR-10 and for 300 epochs on
PTB) and then evaluated on the validation set. For each task, we repeat the experiments for 4 times
with different random seeds, and report the median and the best (per run) validation performance of
the architectures over time. As references, we also report the results (under the same evaluation setup;
with comparable number of parameters) of the best existing cells discovered using RL or evolution,
including NASNet-A (Zoph et al., 2018) (2000 GPU days), AmoebaNet-A (3150 GPU days) (Real
et al., 2018) and ENAS (0.5 GPU day) (Pham et al., 2018b).",0.7814,0.2452,0.5133,896,375,2.389,9a1a747b098dc0a832c939e80a3a6eff,images/2018/arxiv_0000359.png,https://arxiv.org/pdf/1806.09055v2.pdf
arxiv_0000360,Figure 360,scientific_figure,DARTS: Differentiable Architecture Search,arXiv Research Authors,1806.09055v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). DARTS: Differentiable Architecture Search. arXiv:1806.09055v2,Figure 4: Normal cell learned on CIFAR-10.,A diagram of a computer system with.,A detailed layout showing the steps in the sequence of an event in a virtual environment for a web - based.,"A comprehensive technical explanation of the sequence of an application in python with python and python programmings, part 2 of 3, using python to create a script from a sequence of the same sequence of a sequence for a system.",Figure 4: Normal cell learned on CIFAR-10.,0.7321,0.2662,0.4991,421,209,2.014,ee089379ed11030a2f3044febebdc031,images/2018/arxiv_0000360.png,https://arxiv.org/pdf/1806.09055v2.pdf
arxiv_0000361,Figure 361,scientific_figure,DARTS: Differentiable Architecture Search,arXiv Research Authors,1806.09055v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). DARTS: Differentiable Architecture Search. arXiv:1806.09055v2,Figure 6: Recurrent cell learned on PTB.,A diagram of the block diagram for a.,A detailed layout showing the steps of a sequence for a simple algorithm for a random function and a random approach.,"A comprehensive technical explanation of the algorithm to solve the problem for the algorithm in the sequence, a simple model is shown below it is a schematicularized with the following the following two steps of the diagram - 1.",Figure 6: Recurrent cell learned on PTB.,0.75,0.2875,0.5188,340,417,0.815,7a4ce72913caf147581fcfef9e5e793d,images/2018/arxiv_0000361.png,https://arxiv.org/pdf/1806.09055v2.pdf
arxiv_0000362,Figure 362,scientific_figure,DARTS: Differentiable Architecture Search,arXiv Research Authors,1806.09055v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). DARTS: Differentiable Architecture Search. arXiv:1806.09055v2,Figure 5: Reduction cell learned on CIFAR-10.,A diagram of a computer network with.,A detailed layout showing a number of steps that will be used to create a swimming pool in the same way.,"A comprehensive technical explanation of a typical network diagram for a pool pool pool 3, 3, 4, 5, 8, 9, 9 and 10, 1, 0, 2, 2 -, pool 3 - swimming pool 3 representing all details, specifications, and configurations of the system components in full.",Figure 5: Reduction cell learned on CIFAR-10.,0.7214,0.2562,0.4888,420,177,2.373,887e6d08ccaef20d9c9f4d1ab5e40b72,images/2018/arxiv_0000362.png,https://arxiv.org/pdf/1806.09055v2.pdf
arxiv_0000363,Figure 363,scientific_figure,Deep Generative Models with Learnable Knowledge Constraints,arXiv Research Authors,1806.09764v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Generative Models with Learnable Knowledge Constraints. arXiv:1806.09764v2,Human,A diagram of a woman wearing different.,A detailed layout showing different views of a woman with a black top and blue jeans and a dark top.,"A comprehensive technical explanation of the different ways to wear a shirt for women in the sun or in the rain, and how to adjust the neckline and shoulder length on a t - shirt with a woman ' s shoulders,.",Human,0.7071,0.3418,0.5244,768,512,1.5,8673d65b69ec9ed3b336bb08522f277d,images/2018/arxiv_0000363.png,https://arxiv.org/pdf/1806.09764v2.pdf
arxiv_0000364,Figure 364,scientific_figure,Deep Generative Models with Learnable Knowledge Constraints,arXiv Research Authors,1806.09764v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Generative Models with Learnable Knowledge Constraints. arXiv:1806.09764v2,"Figure 1: Two example applications of imposing learnable knowledge constraints on generative
models. Left: Given a person image and a target pose (defined by key points), the goal is to generate
an image of the person under the new pose. The constraint is to force the human parts (e.g., head) of
the generated image to match those of the true target image. Right: Given a text template, the goal is
to generate a complete sentence following the template. The constraint is to force the match between
the infilling content of the generated sentence with the true content. (See sec 5 for more details.)",A diagram of a woman standing in front.,"A detailed layout showing the different types of a woman ' s body and how it works with her that shows key features, attributes, and data points in.","A comprehensive technical explanation of a data processing system for the user of a computer application, including the user and the user ' s workflows of the user area and the system and the application of the process of the software.","Figure 1: Two example applications of imposing learnable knowledge constraints on generative
models. Left: Given a person image and a target pose (defined by key points), the goal is to generate
an image of the person under the new pose. The constraint is to force the human parts (e.g., head) of
the generated image to match those of the true target image. Right: Given a text template, the goal is
to generate a complete sentence following the template. The constraint is to force the match between
the infilling content of the generated sentence with the true content. (See sec 5 for more details.)",0.81,0.2493,0.5297,893,263,3.395,1bc2489e661d0a36848c61eac9cbc8e1,images/2018/arxiv_0000364.png,https://arxiv.org/pdf/1806.09764v2.pdf
arxiv_0000365,Figure 365,scientific_figure,Deep Generative Models with Learnable Knowledge Constraints,arXiv Research Authors,1806.09764v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Generative Models with Learnable Knowledge Constraints. arXiv:1806.09764v2,"Figure 2: Training losses of the three mod-
els. The model with learned constraint con-
verges smoothly as base models.",A diagram of the average and least.,"A detailed layout showing the mean of the data in a computer system, with a line graph and the actual time.","A comprehensive technical explanation of the basic model for the model with fixed constraint and learned constants, based on the data table below, and the data line graph from the data source in the chart below, in the table below.","Figure 2: Training losses of the three mod-
els. The model with learned constraint con-
verges smoothly as base models.",0.7921,0.2397,0.5159,433,271,1.598,20f4ab3abec16ddff639fb44635ef6f0,images/2018/arxiv_0000365.png,https://arxiv.org/pdf/1806.09764v2.pdf
arxiv_0000366,Figure 366,scientific_figure,Deep Generative Models with Learnable Knowledge Constraints,arXiv Research Authors,1806.09764v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Generative Models with Learnable Knowledge Constraints. arXiv:1806.09764v2,"Figure 3: Samples generated by the models in Table 2. The model with learned human part constraint
generates correct poses and preserves human body structure much better.",A diagram of the different styles of.,A detailed layout showing the different types of clothing used by the model and the size of the image.,"A comprehensive technical explanation of the image and how to use it in photoshopping and photo editing techniques for fashion photography, including a video tutor of a woman ' s body image and a woman with a black dress.","Figure 3: Samples generated by the models in Table 2. The model with learned human part constraint
generates correct poses and preserves human body structure much better.",0.7743,0.299,0.5366,914,300,3.047,334caa78e06da329e5ddf6e2f5547b0e,images/2018/arxiv_0000366.png,https://arxiv.org/pdf/1806.09764v2.pdf
arxiv_0000367,Figure 367,scientific_figure,Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning,arXiv Research Authors,1807.06538v6,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning. arXiv:1807.06538v6,"Figure 1: Left: In SMOTE, the minor-class data are randomly selected (denoted as M in the figure).
Further, k-neighbors (3-neighbors in this study) are selected, and pseudo-data are generated at the
interpolating points. Right: Even though the pseudo-data generated by SMOTE cannot go beyond the
determination border, our proposed method can go beyond the border and push it forward near to the
true determination border.",A diagram of the different types of.,"A detailed layout showing the different types of the same data processing system for the same network, with different levels.","A comprehensive technical explanation of the different types of data that are stored in a network of data systems and their uses to track their data for a particular task or task on the task, as well as well, it is now.","Figure 1: Left: In SMOTE, the minor-class data are randomly selected (denoted as M in the figure).
Further, k-neighbors (3-neighbors in this study) are selected, and pseudo-data are generated at the
interpolating points. Right: Even though the pseudo-data generated by SMOTE cannot go beyond the
determination border, our proposed method can go beyond the border and push it forward near to the
true determination border.",0.81,0.2805,0.5453,605,362,1.671,d4dd5329edebbfc8f02f3c9e40d295d6,images/2018/arxiv_0000367.png,https://arxiv.org/pdf/1807.06538v6.pdf
arxiv_0000368,Figure 368,scientific_figure,Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning,arXiv Research Authors,1807.06538v6,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning. arXiv:1807.06538v6,"Figure 2: A schematic of the proposed cavity filling method. Left: Train deep neural networks.
Center: Extract features from a layer, obtain multivariate probability distributions of these features,
and generate pseudo-features of minor classes from the probability distributions and retrain the
following layers. Right: Return the retrained layers to the original one (the final classifier is retrained
and substituted into the experiment.)",A diagram of a classifier and a.,"A detailed layout showing the different features of a classifier and clusterer system, including classes, classes,.","A comprehensive technical explanation of the classifier and classifier feature maps of the system, including the classester, the classifiers, and the classesifiers, and their features and the classier, pngmr.","Figure 2: A schematic of the proposed cavity filling method. Left: Train deep neural networks.
Center: Extract features from a layer, obtain multivariate probability distributions of these features,
and generate pseudo-features of minor classes from the probability distributions and retrain the
following layers. Right: Return the retrained layers to the original one (the final classifier is retrained
and substituted into the experiment.)",0.81,0.2864,0.5482,596,350,1.703,39dedabb27cdf43c7ec65497b77d6bd3,images/2018/arxiv_0000368.png,https://arxiv.org/pdf/1807.06538v6.pdf
arxiv_0000369,Figure 369,scientific_figure,Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning,arXiv Research Authors,1807.06538v6,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning. arXiv:1807.06538v6,"Figure 3: The cavities of the original imbalanced data (left) are filled by pseudo-features in the feature
spaces (right). The pseudo-features are generated using the multivariate probability distributions
obtained from the real features obtained by deep learning.",A diagram of the different features of.,A detailed layout showing the features of the real features in the class model and their features in real features.,"A comprehensive technical explanation of the real features in real features and real features for real features, real features from real features to real features are shown below in this diagram from the table top view in the chart on the table.","Figure 3: The cavities of the original imbalanced data (left) are filled by pseudo-features in the feature
spaces (right). The pseudo-features are generated using the multivariate probability distributions
obtained from the real features obtained by deep learning.",0.87,0.3403,0.6051,516,283,1.823,281e12dcf021cbf0b6eca38dd3d1755f,images/2018/arxiv_0000369.png,https://arxiv.org/pdf/1807.06538v6.pdf
arxiv_0000370,Figure 370,scientific_figure,Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning,arXiv Research Authors,1807.06538v6,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning. arXiv:1807.06538v6,"Figure 4: Comparison of six methods for handling the multi-class imbalanced data (Cifar10): #Minor
denotes the number of minor classes. Top left: accuracy, bottom left: recall, top right: precision, and
bottom right: F1",A diagram of the four stages of a plot.,A detailed layout showing the various types of data that are generated by the same data in different graphs.,"A comprehensive technical explanation of the linear regressions for data visual and modeling of materials in a networked environment, using the data visual tool for data modeling 4cvzmr and r2s and rxs.","Figure 4: Comparison of six methods for handling the multi-class imbalanced data (Cifar10): #Minor
denotes the number of minor classes. Top left: accuracy, bottom left: recall, top right: precision, and
bottom right: F1",0.7957,0.2606,0.5282,3840,1998,1.922,5d79eb8ca2ed52b254d373d22919feea,images/2018/arxiv_0000370.png,https://arxiv.org/pdf/1807.06538v6.pdf
arxiv_0000371,Figure 371,scientific_figure,Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning,arXiv Research Authors,1807.06538v6,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning. arXiv:1807.06538v6,"Figure 5: Comparison of the accuracy, precision, recall, and F1 in only minor classes. Undersampling,
SMOTE, perturbation, and cavity filling exhibit identical scores and overlap; thus, they cannot be
distinguished.",A diagram of the three lines show the.,A detailed layout showing the number of different types of water and their potential to use in the river.,"A comprehensive technical explanation of the data visual of the three graphs in the tableaum data visual tool, including data visual and data visual tools for the data processing tooling tool, and the user interfaces,.","Figure 5: Comparison of the accuracy, precision, recall, and F1 in only minor classes. Undersampling,
SMOTE, perturbation, and cavity filling exhibit identical scores and overlap; thus, they cannot be
distinguished.",0.725,0.209,0.467,3840,1998,1.922,3f35a6b3f2e4051e2cc5ec959f1940f9,images/2018/arxiv_0000371.png,https://arxiv.org/pdf/1807.06538v6.pdf
arxiv_0000372,Figure 372,scientific_figure,Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning,arXiv Research Authors,1807.06538v6,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning. arXiv:1807.06538v6,"Figure 4: Comparison of six methods for handling the multi-class imbalanced data (Cifar10): #Minor
denotes the number of minor classes. Top left: accuracy, bottom left: recall, top right: precision, and
bottom right: F1",A diagram of the different types of.,"A detailed layout showing the different stages of a number of graphs, from one to five in a row that shows key features, attributes, and data points.","A comprehensive technical explanation of the average and expected time frame for the model, and the average time frame, and its corresponding time frame and location, as well as described by the model and the model of the data file size.","Figure 4: Comparison of six methods for handling the multi-class imbalanced data (Cifar10): #Minor
denotes the number of minor classes. Top left: accuracy, bottom left: recall, top right: precision, and
bottom right: F1",0.87,0.2678,0.5689,1224,715,1.712,d1058aa2539e7013ebbb073e7322f0a9,images/2018/arxiv_0000372.png,https://arxiv.org/pdf/1807.06538v6.pdf
arxiv_0000373,Figure 373,scientific_figure,Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning,arXiv Research Authors,1807.06538v6,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Cavity Filling: Pseudo-Feature Generation for Multi-Class Imbalanced Data Problems in Deep Learning. arXiv:1807.06538v6,"Figure 5: Comparison of the accuracy, precision, recall, and F1 in only minor classes. Undersampling,
SMOTE, perturbation, and cavity filling exhibit identical scores and overlap; thus, they cannot be
distinguished.",A diagram of the number of people who.,"A detailed layout showing the different types of data in different graphs, including the same data and the same amount.","A comprehensive technical explanation of the different types of data visual in a software system, including data visual, data visual and data visual graphics, and information graphics, for the web design, and more than data visual visual,.","Figure 5: Comparison of the accuracy, precision, recall, and F1 in only minor classes. Undersampling,
SMOTE, perturbation, and cavity filling exhibit identical scores and overlap; thus, they cannot be
distinguished.",0.81,0.2693,0.5396,1224,431,2.84,ca637575b02bdef44efc4ae215561ea9,images/2018/arxiv_0000373.png,https://arxiv.org/pdf/1807.06538v6.pdf
arxiv_0000374,Figure 374,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 1: The ReferIt task identifies a selected object (in the bounding box) using
a single expression, while in GuessWhat?!, a speaker localizes the object with a
series of yes or no questions.",A diagram of a group of people holding.,"A detailed layout showing a picture of a group of people holding frisbees in the dirt that shows key features, attributes, and data points in detail.","A comprehensive technical explanation of frisbee training for adults and children, including the basics of the game, and how to use it to improve your teamwork on your game development, and play, and maintain your team ' s.","Fig. 1: The ReferIt task identifies a selected object (in the bounding box) using
a single expression, while in GuessWhat?!, a speaker localizes the object with a
series of yes or no questions.",0.81,0.332,0.571,320,240,1.333,f4a35743ae73dbb806cbddfb58eb116c,images/2018/arxiv_0000374.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000375,Figure 375,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 1: The ReferIt task identifies a selected object (in the bounding box) using
a single expression, while in GuessWhat?!, a speaker localizes the object with a
series of yes or no questions.",A diagram of the four people in front.,A detailed layout showing a group of people posing for a picture with a text box containing the words.,A comprehensive technical explanation of the three types of the frisbee game and how it works for kids of all ages to play with them in the field of their age groups or at home or near the woods or near a park.,"Fig. 1: The ReferIt task identifies a selected object (in the bounding box) using
a single expression, while in GuessWhat?!, a speaker localizes the object with a
series of yes or no questions.",0.7143,0.2239,0.4691,773,382,2.024,6d2c8a694194f19992b36730b17032b3,images/2018/arxiv_0000375.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000376,Figure 376,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 2:
The Multi-hop FiLM architecture, illustrating inputs (green), layers
(blue), and activations (purple). In contrast, Single-hop FiLM models predict
FiLM parameters directly from el,T .",A diagram of a computer system with.,"A detailed layout showing the different components of a computer system, including the main processor and the main memory.","A comprehensive technical explanation of the architecture of a computer system, including a memory layer and a large memory layer, including the memory card, and the memory memory card reader, memory, and memory, memory and storage,.","Fig. 2:
The Multi-hop FiLM architecture, illustrating inputs (green), layers
(blue), and activations (purple). In contrast, Single-hop FiLM models predict
FiLM parameters directly from el,T .",0.75,0.2527,0.5013,560,380,1.474,daf27b0030f9904a9637ae048edff537,images/2018/arxiv_0000376.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000377,Figure 377,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 3: Overall model, consisting of a visual pipeline (red and yellow) and linguis-
tic pipeline (blue) and incorporating additional contextual information (green).",A diagram of a cell phone with.,"A detailed layout showing what a computer is connected to and what it is called as a remote system that shows key features, attributes, and data.","A comprehensive technical explanation of the internet phone call system for kids and adults, as well as its features and information, from the internet / t - shirt - shirt / tm / t / s / tf / t representing all details, specifications, and.","Fig. 3: Overall model, consisting of a visual pipeline (red and yellow) and linguis-
tic pipeline (blue) and incorporating additional contextual information (green).",0.81,0.247,0.5285,632,429,1.473,0bb9d8415820a37fe547211f3ce98b72,images/2018/arxiv_0000377.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000378,Figure 378,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 4: Guesser (left) and Oracle (right) attention visualizations for the visual
pipeline which processes the object crop.",A diagram of two men with different.,"A detailed layout showing a blue square and a black square with a red rectangle in it, with a line of words on the left side.","A comprehensive technical explanation of the two men who were in the same team with each other, from the same time they were together, to the next step in the success true story in the game? texting box is true.","Fig. 4: Guesser (left) and Oracle (right) attention visualizations for the visual
pipeline which processes the object crop.",0.75,0.2145,0.4823,457,484,0.944,05a059e67aa437667309f4f6fe803db0,images/2018/arxiv_0000378.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000379,Figure 379,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 4: Guesser (left) and Oracle (right) attention visualizations for the visual
pipeline which processes the object crop.",A diagram of visual reasoning with.,A detailed layout showing a picture of two men in suits and ties with the same name on each of them.,"A comprehensive technical explanation of visual reasoning with multiple - step feature models, 13 / 11 / 12 / 12, and 1 / 3 / 7 / 8 / 8, from the library of the library, through the library library representing all details, specifications, and.","Fig. 4: Guesser (left) and Oracle (right) attention visualizations for the visual
pipeline which processes the object crop.",0.7071,0.2919,0.4995,773,592,1.306,578756eb58fb65912631219c4086c736,images/2018/arxiv_0000379.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000380,Figure 380,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,Attention Visualizations,A diagram of a skier being shown with.,"A detailed layout showing the different levels of a skiers skiing on the slope in the snow that shows key features, attributes, and data points in.","A comprehensive technical explanation of how to use the same level of the camera on skis, skiing, snowboarding and more in the same direction as well as well, the camera, you can see them from the picture below.",Attention Visualizations,0.81,0.3659,0.5879,600,500,1.2,0954e97248d97df1a0d295aa25b9eb24,images/2018/arxiv_0000380.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000381,Figure 381,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 5: The crop pipeline Oracle’s attention over the last question when the model
succeeds.",A diagram of how the player is.,A detailed layout showing the position of a baseball player with the bat in hand and the number of players on the field.,"A comprehensive technical explanation of the baseball player ' s position and how to choose the base, bat, or glove for his swing or the pitch, and the next run, according to avoid that is the runner, what?.","Fig. 5: The crop pipeline Oracle’s attention over the last question when the model
succeeds.",0.81,0.3055,0.5577,600,500,1.2,98f8648dc6792ef15aaf6ef0e22ae3c0,images/2018/arxiv_0000381.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000382,Figure 382,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,NULL,A diagram of what to use the answer.,"A detailed layout showing two cows in a shed with a red and white checkerboard on the wall that shows key features, attributes, and data points in.","A comprehensive technical explanation of the same animal in a shed, with a cow standing and a cow laying in the hay behind it, and a diagram of the different parts on the same level of a third panel of which cow.",NULL,0.81,0.343,0.5765,600,500,1.2,4ed632c5d4821a369e871bd2c10fcd36,images/2018/arxiv_0000382.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000383,Figure 383,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 5: The crop pipeline Oracle’s attention over the last question when the model
succeeds.",A diagram of the various positions of.,"A detailed layout showing the steps of a baseball player ' s swing and the positions for his bat that shows key features, attributes, and data.","A comprehensive technical explanation of the different positions of baseball players in action, from left to right, and on right, a batter, batter, catcher, catcher and umpire, and umpire and umpire on a pitcher, and a horse.","Fig. 5: The crop pipeline Oracle’s attention over the last question when the model
succeeds.",0.87,0.2611,0.5655,773,1299,0.595,2240c7a139b5ac42c8ebc900cc885d39,images/2018/arxiv_0000383.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000384,Figure 384,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 6: The crop pipeline Oracle’s attention over the last question, showing more
advanced reasoning.",A diagram of the progress of a.,"A detailed layout showing the value of a picture with a number of cows in it and a blue sky that shows key features, attributes, and data points in.","A comprehensive technical explanation of the structure and structure of a successful business case for the cow herder ' s success in the field of success and success in a competitive business case, the future or the market, the business case.","Fig. 6: The crop pipeline Oracle’s attention over the last question, showing more
advanced reasoning.",0.81,0.3459,0.5779,600,500,1.2,c5e6de15cdd98417ff8bcc8f6a2e5717,images/2018/arxiv_0000384.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000385,Figure 385,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,NULL,A diagram of an umpire throwing a ball.,"A detailed layout showing the player ' s position and how to use a baseball glove in the game that shows key features, attributes, and data points.","A comprehensive technical explanation of how the game is played by the pitcher and the batter in the next photo, from an interactive video game, to a scoreboard, with a player on one player ' s perspective, or another.",NULL,0.87,0.3394,0.6047,600,500,1.2,28d508f538e77069b3ca11f767c3d1ea,images/2018/arxiv_0000385.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000386,Figure 386,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Visual Reasoning with Multi-hop Feature Modulation
21",A diagram of a skier in mid air with a.,A detailed layout showing the ski jump of a skier in mid air with a level 3 level 5 level 6 level 7 level 8 level 8.,"A comprehensive technical explanation of the skiing slope for skiers in the snow covered mountainside area of the mountain range, showing a level 3 level 3 and a level 5 and a higher level 6, which is below it.","Visual Reasoning with Multi-hop Feature Modulation
21",0.75,0.3048,0.5274,600,500,1.2,7ef1014cfb7eeb13420da312e461afea,images/2018/arxiv_0000386.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000387,Figure 387,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 6: The crop pipeline Oracle’s attention over the last question, showing more
advanced reasoning.",A diagram of the different types of.,A detailed layout showing multiple images of people skiing on the mountain top and snowboarding down the slope.,"A comprehensive technical explanation of the skiing course for beginners and kids, including a guide to learn to ski and snowboard, snowboarding, and snowboarding and snow skiing, skiing, and more thang, and other.","Fig. 6: The crop pipeline Oracle’s attention over the last question, showing more
advanced reasoning.",0.7464,0.2694,0.5079,773,1291,0.599,97279587d62156c934d461c6bee983c8,images/2018/arxiv_0000387.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000388,Figure 388,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"22
F. Strub and M. Seurin and E. Perez and H. Vries et al.",A diagram of a person walking in the.,"A detailed layout showing the amount of snow and how to use it for the camera in an image that shows key features, attributes, and data points in.","A comprehensive technical explanation of the snowboarder ' s winter weather conditions and how to use the system for this photo, which includes the correcting of the correct skis the skiers, and proper skis, the slope.","22
F. Strub and M. Seurin and E. Perez and H. Vries et al.",0.81,0.3467,0.5784,600,500,1.2,644998aa0434ce6119201942c42a7a8c,images/2018/arxiv_0000388.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000389,Figure 389,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 7: The crop pipeline Oracle’s attention over the last question when the model
fails.",A diagram of a bus driving down the.,"A detailed layout showing the bus and the building of a business called answer given no success false,.","A comprehensive technical explanation of a bus on a street in a city with no buses or cars to pass it, and a graphic of a building with a bus driving over the same street name and the same level of the same.","Fig. 7: The crop pipeline Oracle’s attention over the last question when the model
fails.",0.7179,0.3199,0.5189,600,500,1.2,4ad89aae1917d435581aee9d6975676b,images/2018/arxiv_0000389.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000390,Figure 390,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,NULL,A diagram of two bears swimming in a.,A detailed layout showing two bears swimming in the water with a blue line on each side of the image.,"A comprehensive technical explanation of the bear ' s success in the water and how it affects to the environment and how they can help them survive the survivally in the wild world, according by having fun and success in this particular.",NULL,0.7107,0.3471,0.5289,600,500,1.2,297d2530c84735b97b99fc94cdd533a9,images/2018/arxiv_0000390.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000391,Figure 391,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Fig. 7: The crop pipeline Oracle’s attention over the last question when the model
fails.",A diagram of a train and some people.,"A detailed layout showing various types of musical instruments and notes for the music score, including a bear.","A comprehensive technical explanation of the effects of water and air in the environment, including the number of people in the water and the number on the ground above water, and below it in the ground, in the city, below.","Fig. 7: The crop pipeline Oracle’s attention over the last question when the model
fails.",0.8064,0.2058,0.5061,773,1291,0.599,285e74219181079320a1f2d937dd2e27,images/2018/arxiv_0000391.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000392,Figure 392,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"Visual Reasoning with Multi-hop Feature Modulation
23",A diagram of a boat on the water with.,A detailed layout showing the steps to a boat in the water with a line of people standing on the dock.,"A comprehensive technical explanation of the life boat for the ocean rescue team, including the crew ' s safety and maintenance manuals and instructions to the rescue boat ' s crew, from the waterman ' s point of use and the seaweed.","Visual Reasoning with Multi-hop Feature Modulation
23",0.7143,0.2341,0.4742,600,500,1.2,b7bbedbbb5e290134119e3923bba2cce,images/2018/arxiv_0000392.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000393,Figure 393,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,Fig. 8: The crop pipeline Guesser’s attention when the model succeeds.,A diagram of the same stuffed animal.,A detailed layout showing a screen shot of three teddy bears with a green square on the bottom of each one.,A comprehensive technical explanation of a teddy bear and a teddy - bear - and - a - teddy - and a - bear with a square frame at the top of its face and bottom right side of a square image in a square.,Fig. 8: The crop pipeline Guesser’s attention when the model succeeds.,0.7321,0.351,0.5415,600,500,1.2,3b576d6e61f5fc6098332d73f241aeb5,images/2018/arxiv_0000393.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000394,Figure 394,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,NULL,A diagram of the three men are.,A detailed layout showing a photo of four people on a snow covered mountain with a blue line in the middle.,"A comprehensive technical explanation of the success of a snowboarder and his team, including the level 3 level 5 level 6 level 6 levels of difficulty and level 6 layers of the course level 1 level 5 and level 5 levels.",NULL,0.7321,0.3571,0.5446,600,500,1.2,29db239e7b2a2ff7eb316b256e20399a,images/2018/arxiv_0000394.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000395,Figure 395,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,Fig. 8: The crop pipeline Guesser’s attention when the model succeeds.,A diagram of various types of stuffed.,"A detailed layout showing the various stages of a boat and its crew in the snow, with a boat on the water.","A comprehensive technical explanation of visual reasoning with multiple features for children - 2 - 5 years old, 3 - 4 years old and 4 - 6 years old - 9 years old age, and up, and older, and over representing all details, specifications, and.",Fig. 8: The crop pipeline Guesser’s attention when the model succeeds.,0.7286,0.2586,0.4936,699,1295,0.54,c3d4103af37362330bea8cb798707b3a,images/2018/arxiv_0000395.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000396,Figure 396,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,"24
F. Strub and M. Seurin and E. Perez and H. Vries et al.",A diagram of a man walking along a.,"A detailed layout showing the steps to success and the level of success in the photo, with a man carrying luggage.","A comprehensive technical explanation of the layer 3 of a travel guide for travelers in asia and asia, including the success and failure level 3 of the world ' s travel guide, from the top to the next level 5, the world.","24
F. Strub and M. Seurin and E. Perez and H. Vries et al.",0.75,0.3247,0.5373,600,500,1.2,5f3cdf262db0341c54902ebf9c6a2373,images/2018/arxiv_0000396.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000397,Figure 397,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,Fig. 9: The crop pipeline Guesser’s attention when the model fails.,A diagram of the steps in which a.,"A detailed layout showing a few people in different poses and colors, with a blue square on the left side.","A comprehensive technical explanation of the layer 4 layer layer 6 layer 5 layer 6 layers 5 layers and layers of layers of the layers of a layer and a layer 6 component layer of the same layer are shown in this image,.",Fig. 9: The crop pipeline Guesser’s attention when the model fails.,0.7286,0.24,0.4843,600,500,1.2,236cb43db2002819c24cfc9c754b5f71,images/2018/arxiv_0000397.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000398,Figure 398,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,NULL,A diagram of people carrying depicting.,A detailed layout showing the color of two surfers carrying their boards on the beach and a blue sky above them.,"A comprehensive technical explanation of the process for a successful surfboarder ' s success in the ocean and how to use it? - page 3 - layer 3 / layer 2, layer 3, layer 4, layer 5, layer 6, layer representing all details, specifications, and.",NULL,0.75,0.27,0.51,600,500,1.2,09355a17717d1b1b0715eccde653819b,images/2018/arxiv_0000398.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000399,Figure 399,scientific_figure,Visual Reasoning with Multi-hop Feature Modulation,arXiv Research Authors,1808.04446v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Visual Reasoning with Multi-hop Feature Modulation. arXiv:1808.04446v2,Fig. 9: The crop pipeline Guesser’s attention when the model fails.,A diagram of the different stages of a.,A detailed layout showing the steps of the pier and the steps to the beach and the ocean and people walking.,"A comprehensive technical explanation of the steps and steps in a musical instrument, with notes from the same music score, and a photo of a man walking on the beach with a guitar, and on a motorbig,.",Fig. 9: The crop pipeline Guesser’s attention when the model fails.,0.7357,0.2451,0.4904,661,1295,0.51,4906313a0146f645a0d84fff5bd2b82e,images/2018/arxiv_0000399.png,https://arxiv.org/pdf/1808.04446v2.pdf
arxiv_0000400,Figure 400,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 1. High level diagram of the design. G1, G2 are the stage 1 and stage 2 generator. F is the caption generator. We take the
difference between the original caption and the real caption as the loss.",A diagram of a yellow flower with the.,"A detailed layout showing the flow of each flower and its petals in a single layer, with two arrows pointing to the same flower.","A comprehensive technical explanation of the flower arrangement and its meaning in english and spanish language, including the flower ' s size and color scheme, as well as described by the flower petals, as shown below, the diagram,.","Figure 1. High level diagram of the design. G1, G2 are the stage 1 and stage 2 generator. F is the caption generator. We take the
difference between the original caption and the real caption as the loss.",0.81,0.3176,0.5638,569,299,1.903,8b29cbe20c6dce74006560c903e44c94,images/2018/arxiv_0000400.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000401,Figure 401,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 2. Text to Image GAN network. The text embedding and noise is given as input to the first stage. The output of the first stage is
given as input to the next stage that produces higher resolution images. Generators from each stage have corresponding discriminators",A diagram of a block diagram showing.,A detailed layout showing the process of a flower - producing plant with different stages and stages of blooming.,"A comprehensive technical explanation of the process in the process of producing flowers and plants, including the process for making flower bulbs and seeds and flowers to grow them in the next stage 1 stage 2 flowers, 2, 3, 6.","Figure 2. Text to Image GAN network. The text embedding and noise is given as input to the first stage. The output of the first stage is
given as input to the next stage that produces higher resolution images. Generators from each stage have corresponding discriminators",0.75,0.3171,0.5335,747,571,1.308,f797e1514d9e5232b20e406b89aac142,images/2018/arxiv_0000401.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000402,Figure 402,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 1. High level diagram of the design. G1, G2 are the stage 1 and stage 2 generator. F is the caption generator. We take the
difference between the original caption and the real caption as the loss.",A diagram of a computer system with.,"A detailed layout showing the different parts of a computer system, including the cpu and the processor,.","A comprehensive technical explanation of the multi - channel digital camera system for all cameras and video equipment, including a camera head and lens and a tripodic viewfinder, as well as described in this diagram, it is shown.","Figure 1. High level diagram of the design. G1, G2 are the stage 1 and stage 2 generator. F is the caption generator. We take the
difference between the original caption and the real caption as the loss.",0.725,0.2372,0.4811,1224,511,2.395,674b24a4facf5bc178caf31c91b38682,images/2018/arxiv_0000402.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000403,Figure 403,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 2. Text to Image GAN network. The text embedding and noise is given as input to the first stage. The output of the first stage is
given as input to the next stage that produces higher resolution images. Generators from each stage have corresponding discriminators",A diagram of a large amount of.,A detailed layout showing a typical block diagram of a single - channel audio system with two separate channels and two separate speakers.,"A comprehensive technical explanation of the process and its components, including the two main components, are shown below the diagram of the system, the two components, and the diagram below the image, and above it is the diagram,.","Figure 2. Text to Image GAN network. The text embedding and noise is given as input to the first stage. The output of the first stage is
given as input to the next stage that produces higher resolution images. Generators from each stage have corresponding discriminators",0.81,0.2695,0.5398,1224,745,1.643,4b8c864c143c0523bfc48c1d5ed2b17a,images/2018/arxiv_0000403.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000404,Figure 404,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 3. Caption GAN network. The top row shows the caption generator where the LSTM takes CNN features and noise Z as input
and outputs captions. The bottom row shows the discriminator that performs a dot product on the CNN features of the image and the
LSTM output.",A diagram of the process for a flower.,A detailed layout showing the different colors of flowers in a garden and how to use them to make it.,"A comprehensive technical explanation of the basic flow diagram for the flower color scheme, based on the purple flower and the green flower color palettes in the blue and the red one with the black flower is color, the purple,.","Figure 3. Caption GAN network. The top row shows the caption generator where the LSTM takes CNN features and noise Z as input
and outputs captions. The bottom row shows the discriminator that performs a dot product on the CNN features of the image and the
LSTM output.",0.7107,0.2693,0.49,751,636,1.181,21cf143c7c2e4bd666136e42f5f54ba1,images/2018/arxiv_0000404.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000405,Figure 405,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 3. Caption GAN network. The top row shows the caption generator where the LSTM takes CNN features and noise Z as input
and outputs captions. The bottom row shows the discriminator that performs a dot product on the CNN features of the image and the
LSTM output.",A diagram of a flower with four petals.,A detailed layout showing the flow of the process in the process of producing flowers and plants with a single - dimensional design.,"A comprehensive technical explanation of the block diagram for a flower planter ' s garden, including the location of the flowers and the process of growing them to which they will be used in the planters to grow them from the next.","Figure 3. Caption GAN network. The top row shows the caption generator where the LSTM takes CNN features and noise Z as input
and outputs captions. The bottom row shows the discriminator that performs a dot product on the CNN features of the image and the
LSTM output.",0.75,0.2864,0.5182,1224,697,1.756,a323070c0bbeea4117169e4528b2f6a5,images/2018/arxiv_0000405.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000406,Figure 406,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,Figure 4. Stage 1 GAN results which are 64x64 images,A diagram of the different types of.,"A detailed layout showing the different stages of a tulip plant in blooming season, and how it is grown.","A comprehensive technical explanation of the different types of tulips in the garden - page 1, part 2 and 3, with text and photos from the instructions on separate panels in english and a separate background, in separate texts.",Figure 4. Stage 1 GAN results which are 64x64 images,0.7214,0.3103,0.5159,871,274,3.179,04709629c6b25b4aa72095406b5b69a2,images/2018/arxiv_0000406.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000407,Figure 407,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,Figure 5. Stage 2 GAN results which are 128x128 images,A diagram of the process of growing.,"A detailed layout showing the different types of flowers in each of the pictures, including tulips that shows key features, attributes, and data.","A comprehensive technical explanation of the different types of flowers that are in bloom, including the tulips and the irises, and the stamen bulbs, and how they are arranged in this one is being in the same.",Figure 5. Stage 2 GAN results which are 128x128 images,0.81,0.3254,0.5677,871,272,3.202,cc78b4f472e6c2b6b123c346a16d650b,images/2018/arxiv_0000407.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000408,Figure 408,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,Figure 4. Stage 1 GAN results which are 64x64 images,A diagram of the stages of flowers.,A detailed layout showing the different types of flowers that are used in the garden and what they can be seen.,"A comprehensive technical explanation of flower identification and identification in plants with pictures of flowers in different stages of life, including the petals and the stamen on the stem, leaves and the petals, and the stem and the center.",Figure 4. Stage 1 GAN results which are 64x64 images,0.7464,0.3399,0.5432,479,465,1.03,edb272dea0b2ac5db533b65802084ad1,images/2018/arxiv_0000408.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000409,Figure 409,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,Figure 5. Stage 2 GAN results which are 128x128 images,A diagram of different flowers that.,A detailed layout showing the different types of flowers in their blooming stages and the names of each flower.,"A comprehensive technical explanation of the different flowers in each picture, including the petals and the color of the petals, which are different shades, and shapes, and colors, how to see?, from the flower, the same.",Figure 5. Stage 2 GAN results which are 128x128 images,0.7464,0.3269,0.5366,497,352,1.412,b5bb2d9f85685aa66056f5521eb71655,images/2018/arxiv_0000409.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000410,Figure 410,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 6. First row contains ground truth text descriptions, next two rows contain images generated by the GAN trained without cycle
loss and trained with cycle loss respectively. The last row contains captions generated by the captioning network",A diagram of the different types and.,A detailed layout showing the different colors of flowers in different stages of blooming and floweringing time.,"A comprehensive technical explanation of the different varieties of flowers and their names for each flower type in the garden, from the beginning to the end of the year, and beginning of spring, to fall, spring and early summer and autumn.","Figure 6. First row contains ground truth text descriptions, next two rows contain images generated by the GAN trained without cycle
loss and trained with cycle loss respectively. The last row contains captions generated by the captioning network",0.75,0.313,0.5315,870,451,1.929,5fa7a4d555b68590dea5e1bcb82724b7,images/2018/arxiv_0000410.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000411,Figure 411,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 6. First row contains ground truth text descriptions, next two rows contain images generated by the GAN trained without cycle
loss and trained with cycle loss respectively. The last row contains captions generated by the captioning network",A diagram of the different flowers.,A detailed layout showing the different types of flowers in the garden and how they are used to grow.,"A comprehensive technical explanation of the flower garden in spring and summer plants, including the flowers and leaves, and the seeding process, in the picture below the text and pictures below the words on the grid below the picture.","Figure 6. First row contains ground truth text descriptions, next two rows contain images generated by the GAN trained without cycle
loss and trained with cycle loss respectively. The last row contains captions generated by the captioning network",0.7107,0.3195,0.5151,1224,663,1.846,fd91f84abb61da542732ca8bc0b2eb6d,images/2018/arxiv_0000411.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000412,Figure 412,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 7. Figure illustrating mode collapse. The first row contains ground truth text descriptions. The next two rows contain images
generated by the GAN trained without cycle loss and with cycle loss respectively.",A diagram of the different stages of.,A detailed layout showing the stages of flowers and their blooming stages in a garden with multiple photos.,"A comprehensive technical explanation of the different types of flowers and their blooming stages, including the same variety of flowers that are present in each picture, and in the same image, in the following pictures, the flower,.","Figure 7. Figure illustrating mode collapse. The first row contains ground truth text descriptions. The next two rows contain images
generated by the GAN trained without cycle loss and with cycle loss respectively.",0.7321,0.3243,0.5282,869,396,2.194,c65c953121bd17d25aa72e57e7dc5b14,images/2018/arxiv_0000412.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000413,Figure 413,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 8. Results of GAN trained with cycle loss with image cap-
tioning network’s weights frozen",A diagram of the four different types.,"A detailed layout showing the different types of flowers and their names, including the flower, the petals, and the petals.","A comprehensive technical explanation of the types of flowers in the garden and their uses in the field, including the identification of each plant and its blooming stage, and how it is used to grow, and what they are used.","Figure 8. Results of GAN trained with cycle loss with image cap-
tioning network’s weights frozen",0.75,0.3151,0.5325,439,409,1.073,1ee4855e76fd69a46a65aa531b549104,images/2018/arxiv_0000413.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000414,Figure 414,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 7. Figure illustrating mode collapse. The first row contains ground truth text descriptions. The next two rows contain images
generated by the GAN trained without cycle loss and with cycle loss respectively.",A diagram of the different flowers in.,"A detailed layout showing the different stages of flowering in the garden, with pictures of flowers and text.","A comprehensive technical explanation of flowers and their meanings in the classroom of children and adults, complete with pictures and texting, and instructions for each subject in english or spanish, in the following the following notes and in.","Figure 7. Figure illustrating mode collapse. The first row contains ground truth text descriptions. The next two rows contain images
generated by the GAN trained without cycle loss and with cycle loss respectively.",0.7393,0.3578,0.5485,1224,602,2.033,a00788cd2e174a425c605bbc622cc7f1,images/2018/arxiv_0000414.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000415,Figure 415,scientific_figure,Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks,arXiv Research Authors,1808.04538v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Text-to-Image-to-Text Translation using Cycle Consistent Adversarial Networks. arXiv:1808.04538v1,"Figure 8. Results of GAN trained with cycle loss with image cap-
tioning network’s weights frozen",A diagram of flowering plants in the.,A detailed layout showing the different types of flowers and their meaning and uses for each type of flower.,A comprehensive technical explanation of the flowering patterns and their characteristics for a garden plant or flower bed and how to use them to grow flowers? - page 2 - 3 - page - 1 - 2 - from the wildflowers.,"Figure 8. Results of GAN trained with cycle loss with image cap-
tioning network’s weights frozen",0.7357,0.3336,0.5347,549,520,1.056,7e2193bba018dca4584fe6f3069168ff,images/2018/arxiv_0000415.png,https://arxiv.org/pdf/1808.04538v1.pdf
arxiv_0000416,Figure 416,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 1: Graphical representation of the proposed model for
classification.",A diagram of a block diagram with two.,A detailed layout showing the basic components of a multi - level system for a single - phase circuit.,"A comprehensive technical explanation of the basic system for the use of the computer and other electronic devices, including the two - way switcher and the two switches on each - way switches, the same - on the - off -.","Fig. 1: Graphical representation of the proposed model for
classification.",0.7143,0.2494,0.4819,1043,479,2.177,3676f6b9073b682ebf9a1584b3c2109c,images/2018/arxiv_0000416.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000417,Figure 417,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 1: Graphical representation of the proposed model for
classification.",A diagram of an electronic device with.,A detailed layout showing the three phases of a transformer and the three different phases of an invertible.,"A comprehensive technical explanation of the ieee transaction system, based on a single - line diagram of an internet network and the data processing process in a single device, with multiple devices, together, in a series, each, one.","Fig. 1: Graphical representation of the proposed model for
classification.",0.7357,0.2239,0.4798,584,322,1.814,0f22fdebaf8787642e1d79bb5c26c32c,images/2018/arxiv_0000417.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000418,Figure 418,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 2: Example of 4 different annotators (represented by
different colours) with different biases and precisions.",A diagram of the relationship between.,A detailed layout showing the difference of true rating and the value of the corresponding rate of the rate.,"A comprehensive technical explanation of the true rating curve for a graph of interest in a course of a particular time period of study, including time and date of study time of the next stages of the course of course of study.","Fig. 2: Example of 4 different annotators (represented by
different colours) with different biases and precisions.",0.7357,0.306,0.5209,800,600,1.333,afda1b6c500a2a2d1b07f5a6b68e5304,images/2018/arxiv_0000418.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000419,Figure 419,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 3: Graphical representation of the proposed model for
regression. own personal bias br and precision pr (inverse variance),
and assuming a Gaussian noise model for the annotators’
answers, we have that yd,r ∼N(yd,r|xd + br, 1/pr).
(9) This approach is therefore more powerful than previous
works [22], [24], where a single precision parameter was
used to model the annotators’ expertise. Fig. 2 illustrates this
intuition for 4 annotators, represented by different colors.
The “green annotator” is the best one, since he is right
on the target and his answers vary very little (low bias,
high precision). The “yellow annotator” has a low bias,
but his answers are very uncertain, as they can vary a
lot. Contrarily, the “blue annotator” is very precise, but
consistently over-estimates the true target (high bias, high
precision). Finally, the “red annotator” corresponds to the
worst kind of annotator (with high bias and low precision).
Having specified a model for annotators answers given
the true targets, the only thing left is to do is to specify
a model of the latent true targets xd given the empirical
topic mixture distributions ¯zd. For this, we shall keep things
simple and assume a linear model as in sLDA [7]. The
generative process of the proposed model for continuous
target variables can then be summarized as follows:
1) For each annotator r
a) For each class c
i) Draw reliability parameter πr
c|ω ∼Dir(ω)
2) For each topic k
a) Draw topic distribution βk|τ ∼Dir(τ)
3) For each document d
a) Draw topic proportions θd|α ∼Dir(α)
b) For the nth word
i) Draw topic assignment zd
n|θd ∼Mult(θd)
ii) Draw word wd
n|zd
n, β ∼Mult(βzdn) c) Draw latent (true) target xd|zd, η, σ ∼N(ηT ¯zd, σ2)
d) For each annotator r ∈Rd
i) Draw answer yd,r|xd, br, pr ∼N(xd + br, 1/pr)
Fig. 3 shows a graphical representation of the proposed
model. 4.2
Approximate inference The goal of inference is to compute the posterior distribution
of the per-document topic proportions θd, the per-word
topic assignments zd
n, the per-topic distribution over words
βk and the per-document latent true targets xd. As we did",A diagram of a block diagram with two.,"A detailed layout showing the basic model of an array of data sources and their corresponding output types, including.","A comprehensive technical explanation of the data flow diagram for the system, with diagrams and instructions to use it to make sure that the data flows in the system is accurate and how it workspaces are done for each way.","Fig. 3: Graphical representation of the proposed model for
regression. own personal bias br and precision pr (inverse variance),
and assuming a Gaussian noise model for the annotators’
answers, we have that yd,r ∼N(yd,r|xd + br, 1/pr).
(9) This approach is therefore more powerful than previous
works [22], [24], where a single precision parameter was
used to model the annotators’ expertise. Fig. 2 illustrates this
intuition for 4 annotators, represented by different colors.
The “green annotator” is the best one, since he is right
on the target and his answers vary very little (low bias,
high precision). The “yellow annotator” has a low bias,
but his answers are very uncertain, as they can vary a
lot. Contrarily, the “blue annotator” is very precise, but
consistently over-estimates the true target (high bias, high
precision). Finally, the “red annotator” corresponds to the
worst kind of annotator (with high bias and low precision).
Having specified a model for annotators answers given
the true targets, the only thing left is to do is to specify
a model of the latent true targets xd given the empirical
topic mixture distributions ¯zd. For this, we shall keep things
simple and assume a linear model as in sLDA [7]. The
generative process of the proposed model for continuous
target variables can then be summarized as follows:
1) For each annotator r
a) For each class c
i) Draw reliability parameter πr
c|ω ∼Dir(ω)
2) For each topic k
a) Draw topic distribution βk|τ ∼Dir(τ)
3) For each document d
a) Draw topic proportions θd|α ∼Dir(α)
b) For the nth word
i) Draw topic assignment zd
n|θd ∼Mult(θd)
ii) Draw word wd
n|zd
n, β ∼Mult(βzdn) c) Draw latent (true) target xd|zd, η, σ ∼N(ηT ¯zd, σ2)
d) For each annotator r ∈Rd
i) Draw answer yd,r|xd, br, pr ∼N(xd + br, 1/pr)
Fig. 3 shows a graphical representation of the proposed
model. 4.2
Approximate inference The goal of inference is to compute the posterior distribution
of the per-document topic proportions θd, the per-word
topic assignments zd
n, the per-topic distribution over words
βk and the per-document latent true targets xd. As we did",0.81,0.2829,0.5464,1010,471,2.144,81b150fd00208a52fb0b011607052c29,images/2018/arxiv_0000419.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000420,Figure 420,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 3: Graphical representation of the proposed model for
regression.",A diagram of the two groups of three.,"A detailed layout showing the circuit and components for a voltage - in - voltage generator, including the two switches.","A comprehensive technical explanation of the typical circuit diagram for an acr circuit using vx - xl and vxx - xxxxx, vxvx, the following versions are available with a / vx representing all details, specifications, and configurations of the system.","Fig. 3: Graphical representation of the proposed model for
regression.",0.75,0.2215,0.4858,584,323,1.808,1bf56de108b43711cded05c364b28464,images/2018/arxiv_0000420.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000421,Figure 421,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 2: Example of 4 different annotators (represented by
different colours) with different biases and precisions.",A diagram of the different functions.,"A detailed layout showing the flow of data from a data center to an internet network, including the following data.","A comprehensive technical explanation of the use of artificial intelligence in real - time data processing and design, using ibm ' s new ais platform, the intelliveerraxs and intell software, v4r representing all details, specifications, and.","Fig. 2: Example of 4 different annotators (represented by
different colours) with different biases and precisions.",0.81,0.2077,0.5089,584,431,1.355,7b18e3a1d38570abd29a09a005da82c9,images/2018/arxiv_0000421.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000422,Figure 422,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 5: Comparison of the log marginal likelihood between
the batch and the stochastic variational inference (svi) algo-
rithms on the 20-Newsgroups corpus.",A diagram of a plot that shows the.,"A detailed layout showing the distribution of the mass and the mass / dla data for each sample that shows key features, attributes, and data points.","A comprehensive technical explanation of the mean of the mass - dm in the human body and the number of hours for each of the samples that are present in the sample, in the study, including the sample results is not.","Fig. 5: Comparison of the log marginal likelihood between
the batch and the stochastic variational inference (svi) algo-
rithms on the 20-Newsgroups corpus.",0.87,0.3071,0.5886,543,386,1.407,c6e1d0d5ad1c92efc0534011f183e8d8,images/2018/arxiv_0000422.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000423,Figure 423,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 4: Average testset accuracy (over 5 runs; ± stddev.) of
the different approaches on the 20-Newsgroups data.",A diagram of the line graph shows that.,A detailed layout showing the number of different types of radiations in the atmosphere and the number and range.,"A comprehensive technical explanation of the h - 2b and h - 3b h - 4b h2b h3b h5b h6b h4c h5h h1c0c h2 representing all details, specifications, and configurations of the system components in full context with annotated elements.","Fig. 4: Average testset accuracy (over 5 runs; ± stddev.) of
the different approaches on the 20-Newsgroups data.",0.75,0.2509,0.5005,565,503,1.123,51e30a0d5ae93d222fd7968a17ed7a74,images/2018/arxiv_0000423.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000424,Figure 424,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 6: Boxplot of the number of answers per worker (a) and
their respective accuracies (b) for the Reuters dataset.",A diagram of the results of two.,"A detailed layout showing the number of possible locations for different locations in the area of the area, and where.","A comprehensive technical explanation of the three different models for a real time machine - driven data system - image 1 of 2 - - © 2010, university ofsquareo / edi - eduv, university - ed representing all details, specifications, and.","Fig. 6: Boxplot of the number of answers per worker (a) and
their respective accuracies (b) for the Reuters dataset.",0.75,0.2216,0.4858,553,311,1.778,d76be0f1ddc3e2cf265290bb9ba75c02,images/2018/arxiv_0000424.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000425,Figure 425,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 7: Average testset accuracy (over 30 runs; ± stddev.) of
the different approaches on the Reuters data.",A diagram of the different types of.,A detailed layout showing the distribution of the data for the three transactions on pattern analysis and machine intelligence.,"A comprehensive technical explanation of the three transfusions on pattern analysis and machine intelligence, based on data driven data processing, and machine - driven design, c - driven modeling, cx - based data - driven,.","Fig. 7: Average testset accuracy (over 30 runs; ± stddev.) of
the different approaches on the Reuters data.",0.81,0.2792,0.5446,565,503,1.123,3cd87af2c5718ccdaa90d36211d373a9,images/2018/arxiv_0000425.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000426,Figure 426,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 9: Average testset accuracy (over 30 runs; ± stddev.) of
the different approaches on the LabelMe data.",A diagram of the results of a large.,A detailed layout showing the mean of the initial and initial of the sample results for each of the two samples.,"A comprehensive technical explanation of the various phosphers and their characteristics for each of the three phases of the development of the x - ray system, including the y - ray, y - rx, x - y - y.","Fig. 9: Average testset accuracy (over 30 runs; ± stddev.) of
the different approaches on the LabelMe data.",0.81,0.2428,0.5264,550,503,1.093,51bcf64296c3c45fe3d70d9e5a947a95,images/2018/arxiv_0000426.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000427,Figure 427,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 8: Boxplot of the number of answers per worker (a) and
their respective accuracies (b) for the LabelMe dataset.",A diagram of a square with lines that.,"A detailed layout showing the different types of the square and rectangled boxes used in this work that shows key features, attributes, and data.","A comprehensive technical explanation of the data visual model for a scatter plot of two rectangles, with a line of measurement between them and a box of 20 x - 1, with 2, 0, 1, 3 representing all details, specifications, and configurations of the.","Fig. 8: Boxplot of the number of answers per worker (a) and
their respective accuracies (b) for the LabelMe dataset.",0.81,0.2631,0.5366,522,309,1.689,b7c4e0b3ad7f77f3a55858b2821c3d58,images/2018/arxiv_0000427.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000428,Figure 428,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,Fig. 10: True vs. estimated confusion matrix (cm) of 6 different workers of the Reuters-21578 dataset.,A diagram of the distribution of the.,"A detailed layout showing the number of individuals with different type of disease in the past years, and how they.","A comprehensive technical explanation of the mean of a data - driven data processing system for data processing and processing of data, with data visual and data processing, from data to data processing to data retrieval, and visual, and storage.",Fig. 10: True vs. estimated confusion matrix (cm) of 6 different workers of the Reuters-21578 dataset.,0.75,0.2488,0.4994,1096,463,2.367,10bad845056c0da19621331bab52bd78,images/2018/arxiv_0000428.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000429,Figure 429,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,and their respective biases,A diagram of the city skyline with a.,"A detailed layout showing the city skyline and surrounding buildings, including skyscrapers and a clock tower in the distance.","A comprehensive technical explanation of the city skyline from the top of a building in the foreground, including the tall buildings and the trees that are in the center of the fore ground and to the left of the right side of the picture.",and their respective biases,0.75,0.2861,0.5181,256,256,1.0,8f6aeb31a2734068929df36017d2af5e,images/2018/arxiv_0000429.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000430,Figure 430,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 11: Average testset R2 (over 30 runs; ± stddev.) of the
different approaches on the we8there data.",A diagram of the entrance to the.,"A detailed layout showing the entrance to a building with a clock on it ' s wall and a large door that shows key features, attributes, and data.","A comprehensive technical explanation of the entrance to the new york state college building, new york city, ny, usa, 1971 - 1982 by david mc mcie, via wikipediae, 2013, via wikimer, via commons representing all details, specifications, and.","Fig. 11: Average testset R2 (over 30 runs; ± stddev.) of the
different approaches on the we8there data.",0.81,0.2585,0.5343,256,256,1.0,4ba2d8ebe0df48a000a73677af7e9694,images/2018/arxiv_0000430.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000431,Figure 431,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 11 shows the results obtained for different numbers
of topics. Do to the stochastic nature of both the annotators
simulation procedure and the initialization of the variational
Bayesian EM algorithm, we repeated each experiment 30
times and report the average R2 obtained with the corre-
sponding standard deviation. Since the regression datasets
that are considered in this article are not large enough to
justify the use of a stochastic variational inference (svi) algo-
rithm, we only made experiments using the batch algorithm
developed in Section 4.2. The results obtained clearly show
the improved performance of MA-sLDAr over the other
methods.",A diagram of a sidewalk near some.,A detailed layout showing the grass and the surrounding buildings outside of an apartment complex with a walkway in between.,"A comprehensive technical explanation of apartment building design and construction in the united states, including residential buildings and apartments, with multiple balconies and balconings and balks on the front of each building, and sides, as.","Fig. 11 shows the results obtained for different numbers
of topics. Do to the stochastic nature of both the annotators
simulation procedure and the initialization of the variational
Bayesian EM algorithm, we repeated each experiment 30
times and report the average R2 obtained with the corre-
sponding standard deviation. Since the regression datasets
that are considered in this article are not large enough to
justify the use of a stochastic variational inference (svi) algo-
rithm, we only made experiments using the batch algorithm
developed in Section 4.2. The results obtained clearly show
the improved performance of MA-sLDAr over the other
methods.",0.75,0.3187,0.5343,256,256,1.0,88cc547cd224b67b503115355ac7ebe5,images/2018/arxiv_0000431.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000432,Figure 432,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 11: Average testset R2 (over 30 runs; ± stddev.) of the
different approaches on the we8there data.",A diagram of a plot with multiple.,"A detailed layout showing the size and characteristics of the lvx - x and vx - vxx that shows key features, attributes, and data points in detail.","A comprehensive technical explanation of the lmrvx - x and mzx - txxxx - 2 line graphs for the current vlc range of the two different systems, including the same timeframes representing all details, specifications, and configurations of the system.","Fig. 11: Average testset R2 (over 30 runs; ± stddev.) of the
different approaches on the we8there data.",0.81,0.2242,0.5171,550,498,1.104,c5fd6f8e93ccae77f9bb4fc252cddea9,images/2018/arxiv_0000432.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000433,Figure 433,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 12: Boxplot of the number of answers per worker (a) and
their respective biases (b) and variances (c) for the movie
reviews dataset.",A diagram of the two different types.,"A detailed layout showing the different types of plots and numbers of samples of various materials, including a number.","A comprehensive technical explanation of the different types of the data processing system for the software for windows 7 and 8 1, including the data visual and visual visual control system for each device, from the following stage 3 / 4 /.","Fig. 12: Boxplot of the number of answers per worker (a) and
their respective biases (b) and variances (c) for the movie
reviews dataset.",0.81,0.2781,0.5441,553,330,1.676,d8357290ae44df98688ff59175a178a5,images/2018/arxiv_0000433.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000434,Figure 434,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 14: True vs. predicted biases and precisions of 10 ran-
dom workers of the movie reviews dataset.",A diagram of the number of people in.,"A detailed layout showing the different types of the number of patients in each type of patient, including the number and type of patients.","A comprehensive technical explanation of the number and type of parameters for a computer system, including the number of parameters and the number for a software machine that is used to monitor the device in the system on the device or device,.","Fig. 14: True vs. predicted biases and precisions of 10 ran-
dom workers of the movie reviews dataset.",0.75,0.1946,0.4723,492,362,1.359,3915dac3478914e67f632ead39e3a533,images/2018/arxiv_0000434.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000435,Figure 435,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Fig. 13: Average testset R2 (over 30 runs; ± stddev.) of the
different approaches on the movie reviews data.",A diagram of a plot with a line graph.,"A detailed layout showing the mean of the mean for a single - phase linear plot, and a line graph that shows key features, attributes, and data.","A comprehensive technical explanation of the lga - r mean for a single - cell cell phone system, with data and time in the form of a line graphing function chart of the chart to the data for each cell line.","Fig. 13: Average testset R2 (over 30 runs; ± stddev.) of the
different approaches on the movie reviews data.",0.93,0.2894,0.6097,530,498,1.064,5bd492ec842e116d4ee9f0631a31abd4,images/2018/arxiv_0000435.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000436,Figure 436,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"[6]
E. Airoldi, D. Blei, S. Fienberg, and E. Xing, “Combining stochastic
block models and mixed membership for statistical network anal-
ysis,” in Statistical Network Analysis.
Springer, 2007, pp. 57–74.
[7]
J. Mcauliffe and D. Blei, “Supervised topic models,” in Advances in
neural information processing systems, 2008, pp. 121–128.
[8]
J. Zhu, A. Ahmed, and E. Xing, “Medlda: Maximum margin
supervised topic models,” J. Mach. Learn. Res., vol. 13, no. 1, pp.
2237–2278, 2012.
[9]
R. Snow, B. O’Connor, D. Jurafsky, and A. Ng, “Cheap and fast
- but is it good?: Evaluating non-expert annotations for natural
language tasks,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2008, pp. 254–263.
[10] F. Rodrigues, F. Pereira, and B. Ribeiro, “Learning from multiple
annotators: distinguishing good from random labelers,” Pattern
Recognition Letters, pp. 1428–1436, 2013.
[11] P. Welinder, S. Branson, P. Perona, and S. Belongie, “The multidi-
mensional wisdom of crowds,” in Advances in neural information
processing systems, 2010, pp. 2424–2432.
[12] Y. Yan, R. Rosales, G. Fung, R. Subramanian, and J. Dy, “Learning
from multiple annotators with varying expertise,” Mach. Learn.,
vol. 95, no. 3, pp. 291–327, 2014.
[13] F. Rodrigues, M. Lourenc¸o, F. Pereira, and B. Ribeiro, “Learning
supervised topic models from crowds,” in Proc. of the Third AAAI
Conf. on Human Computation and Crowdsourcing, 2015, pp. 160–168.
[14] M. Hoffman, D. Blei, C. Wang, and J. Paisley, “Stochastic varia-
tional inference,” J. Mach. Learn. Res., vol. 14, pp. 1303–1347, 2013.
[15] S. Lacoste-Julien, F. Sha, and M. Jordan, “Disclda: Discriminative
learning for dimensionality reduction and classification,” in Ad-
vances in neural information processing systems, 2009, pp. 897–904.
[16] D. Ramage, D. Hall, R. Nallapati, and C. Manning, “Labeled LDA:
A supervised topic model for credit attribution in multi-labeled
corpora,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2009, pp. 248–256.
[17] D. Mimno and A. McCallum, “Topic models conditioned on
arbitrary features with dirichlet-multinomial regression,” Proc. of
The 24th Conf. on Uncertainty in Artificial Intelligence, 2008.
[18] M. Rabinovich and D. Blei, “The inverse regression topic model,”
in Proc. of The 31st Int. Conf. on Machine Learning, 2014, pp. 199–207.
[19] M. Taddy, “Multinomial inverse regression for text analysis,” J. of
the American Statistical Assoc., vol. 108, no. 503, pp. 755–770, 2013.
[20] A. P. Dawid and A. M. Skene, “Maximum likelihood estimation
of observer error-rates using the EM algorithm,” J. of the Royal
Statistical Society. Series C, vol. 28, no. 1, pp. 20–28, 1979.
[21] P. Smyth, U. Fayyad, M. Burl, P. Perona, and P. Baldi, “Inferring
ground truth from subjective labelling of venus images,” in Ad-
vances in Neural Information Processing Systems, 1995, pp. 1085–1092.
[22] V. Raykar, S. Yu, L. Zhao, G. Valadez, C. Florin, L. Bogoni, and
L. Moy, “Learning from Crowds,” J. Mach. Learn. Res, pp. 1297–
1322, 2010.
[23] F. Rodrigues, F. Pereira, and B. Ribeiro, “Gaussian process classi-
fication and active learning with multiple annotators,” in Proc. of
the 31st Int. Conf. on Machine Learning, 2014, pp. 433–441.
[24] P. Groot, A. Birlutiu, and T. Heskes, “Learning from multiple
annotators with Gaussian processes,” in Proc. of the Int. Conf. on
Artificial Neural Networks, vol. 6792, 2011, pp. 159–164.
[25] Y. Yan, R. Rosales, G. Fung, M. Schmidt, G. Valadez, L. Bogoni,
L. Moy, and J. Dy, “Modeling annotator expertise: Learning when
everybody knows a bit of something,” J. Mach. Learn. Res, vol. 9,
pp. 932–939, 2010.
[26] M. I. Jordan, Z. Ghahramani, T. Jaakkola, and L. Saul, “An in-
troduction to variational methods for graphical models,” Mach.
Learn., vol. 37, no. 2, pp. 183–233, 1999.
[27] J. Chuang, S. Gupta, C. Manning, and J. Heer, “Topic model
diagnostics: Assessing domain relevance via topical alignment.”
in Proc. of the 30th Int. Conf. on Machine Learning, 2013, pp. 612–620.
[28] J. Nocedal and S. Wright, Numerical Optimization. World Sc., 2006.
[29] H. Robbins and S. Monro, “A stochastic approximation method,”
Ann. Math. Statist., vol. 22, no. 3, pp. 400–407, 09 1951.
[30] K. Lang, “Newsweeder: Learning to filter netnews,” in Proc. of the
Int. Conf. on Machine Learning, 1995, pp. 331–339.
[31] D. Lewis, “Reuters-21578 text categorization test collection. distri-
bution 1.0. readme file (version 1.2),” 1997. [Online]. Available:
www.daviddlewis.com/resources/testcollections/reuters21578/
[32] B. Russell, A. Torralba, K. Murphy, and W. Freeman, “Labelme:
a database and web-based tool for image annotation,” Int. J. of
Computer Vision, vol. 77, no. 1-3, pp. 157–173, 2008.",A diagram of the head and shoulders of.,"A detailed layout showing the man ' s face and glasses, but he is not happy with his tie that shows key features, attributes, and data points in.","A comprehensive technical explanation of this is a black and white photo of a man with glasses and a sweater jacket on, his face slightly to the side, looking at the camera, in front, and behind him, he is a plain.","[6]
E. Airoldi, D. Blei, S. Fienberg, and E. Xing, “Combining stochastic
block models and mixed membership for statistical network anal-
ysis,” in Statistical Network Analysis.
Springer, 2007, pp. 57–74.
[7]
J. Mcauliffe and D. Blei, “Supervised topic models,” in Advances in
neural information processing systems, 2008, pp. 121–128.
[8]
J. Zhu, A. Ahmed, and E. Xing, “Medlda: Maximum margin
supervised topic models,” J. Mach. Learn. Res., vol. 13, no. 1, pp.
2237–2278, 2012.
[9]
R. Snow, B. O’Connor, D. Jurafsky, and A. Ng, “Cheap and fast
- but is it good?: Evaluating non-expert annotations for natural
language tasks,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2008, pp. 254–263.
[10] F. Rodrigues, F. Pereira, and B. Ribeiro, “Learning from multiple
annotators: distinguishing good from random labelers,” Pattern
Recognition Letters, pp. 1428–1436, 2013.
[11] P. Welinder, S. Branson, P. Perona, and S. Belongie, “The multidi-
mensional wisdom of crowds,” in Advances in neural information
processing systems, 2010, pp. 2424–2432.
[12] Y. Yan, R. Rosales, G. Fung, R. Subramanian, and J. Dy, “Learning
from multiple annotators with varying expertise,” Mach. Learn.,
vol. 95, no. 3, pp. 291–327, 2014.
[13] F. Rodrigues, M. Lourenc¸o, F. Pereira, and B. Ribeiro, “Learning
supervised topic models from crowds,” in Proc. of the Third AAAI
Conf. on Human Computation and Crowdsourcing, 2015, pp. 160–168.
[14] M. Hoffman, D. Blei, C. Wang, and J. Paisley, “Stochastic varia-
tional inference,” J. Mach. Learn. Res., vol. 14, pp. 1303–1347, 2013.
[15] S. Lacoste-Julien, F. Sha, and M. Jordan, “Disclda: Discriminative
learning for dimensionality reduction and classification,” in Ad-
vances in neural information processing systems, 2009, pp. 897–904.
[16] D. Ramage, D. Hall, R. Nallapati, and C. Manning, “Labeled LDA:
A supervised topic model for credit attribution in multi-labeled
corpora,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2009, pp. 248–256.
[17] D. Mimno and A. McCallum, “Topic models conditioned on
arbitrary features with dirichlet-multinomial regression,” Proc. of
The 24th Conf. on Uncertainty in Artificial Intelligence, 2008.
[18] M. Rabinovich and D. Blei, “The inverse regression topic model,”
in Proc. of The 31st Int. Conf. on Machine Learning, 2014, pp. 199–207.
[19] M. Taddy, “Multinomial inverse regression for text analysis,” J. of
the American Statistical Assoc., vol. 108, no. 503, pp. 755–770, 2013.
[20] A. P. Dawid and A. M. Skene, “Maximum likelihood estimation
of observer error-rates using the EM algorithm,” J. of the Royal
Statistical Society. Series C, vol. 28, no. 1, pp. 20–28, 1979.
[21] P. Smyth, U. Fayyad, M. Burl, P. Perona, and P. Baldi, “Inferring
ground truth from subjective labelling of venus images,” in Ad-
vances in Neural Information Processing Systems, 1995, pp. 1085–1092.
[22] V. Raykar, S. Yu, L. Zhao, G. Valadez, C. Florin, L. Bogoni, and
L. Moy, “Learning from Crowds,” J. Mach. Learn. Res, pp. 1297–
1322, 2010.
[23] F. Rodrigues, F. Pereira, and B. Ribeiro, “Gaussian process classi-
fication and active learning with multiple annotators,” in Proc. of
the 31st Int. Conf. on Machine Learning, 2014, pp. 433–441.
[24] P. Groot, A. Birlutiu, and T. Heskes, “Learning from multiple
annotators with Gaussian processes,” in Proc. of the Int. Conf. on
Artificial Neural Networks, vol. 6792, 2011, pp. 159–164.
[25] Y. Yan, R. Rosales, G. Fung, M. Schmidt, G. Valadez, L. Bogoni,
L. Moy, and J. Dy, “Modeling annotator expertise: Learning when
everybody knows a bit of something,” J. Mach. Learn. Res, vol. 9,
pp. 932–939, 2010.
[26] M. I. Jordan, Z. Ghahramani, T. Jaakkola, and L. Saul, “An in-
troduction to variational methods for graphical models,” Mach.
Learn., vol. 37, no. 2, pp. 183–233, 1999.
[27] J. Chuang, S. Gupta, C. Manning, and J. Heer, “Topic model
diagnostics: Assessing domain relevance via topical alignment.”
in Proc. of the 30th Int. Conf. on Machine Learning, 2013, pp. 612–620.
[28] J. Nocedal and S. Wright, Numerical Optimization. World Sc., 2006.
[29] H. Robbins and S. Monro, “A stochastic approximation method,”
Ann. Math. Statist., vol. 22, no. 3, pp. 400–407, 09 1951.
[30] K. Lang, “Newsweeder: Learning to filter netnews,” in Proc. of the
Int. Conf. on Machine Learning, 1995, pp. 331–339.
[31] D. Lewis, “Reuters-21578 text categorization test collection. distri-
bution 1.0. readme file (version 1.2),” 1997. [Online]. Available:
www.daviddlewis.com/resources/testcollections/reuters21578/
[32] B. Russell, A. Torralba, K. Murphy, and W. Freeman, “Labelme:
a database and web-based tool for image annotation,” Int. J. of
Computer Vision, vol. 77, no. 1-3, pp. 157–173, 2008.",0.81,0.2402,0.5251,1250,1562,0.8,5077447447df741e34c448bdb1a208a4,images/2018/arxiv_0000436.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000437,Figure 437,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"Mariana Lourenc¸o has a MSc degree in In-
formatics Engineering from University of Coim-
bra, Portugal. Her thesis presented a supervised
topic model that is able to learn from crowds and
she took part in a research project whose pri-
mary objective was to exploit online information
about public events to build predictive models of
flows of people in the city. Her main research in-
terests are machine learning, pattern recognition
and natural language processing.",A diagram of a woman sitting on a.,"A detailed layout showing the front of a woman ' s face and shoulder length hair, smiling, in a car.","A comprehensive technical explanation of the missing girl ' s missing picture, taken in the car with her dog on the back seat of the vehicle, is very close up to her face and smiling now - sized, she.","Mariana Lourenc¸o has a MSc degree in In-
formatics Engineering from University of Coim-
bra, Portugal. Her thesis presented a supervised
topic model that is able to learn from crowds and
she took part in a research project whose pri-
mary objective was to exploit online information
about public events to build predictive models of
flows of people in the city. Her main research in-
terests are machine learning, pattern recognition
and natural language processing.",0.7071,0.2492,0.4781,196,246,0.797,c2414829f274cd0a61d30bbf06389927,images/2018/arxiv_0000437.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000438,Figure 438,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"[6]
E. Airoldi, D. Blei, S. Fienberg, and E. Xing, “Combining stochastic
block models and mixed membership for statistical network anal-
ysis,” in Statistical Network Analysis.
Springer, 2007, pp. 57–74.
[7]
J. Mcauliffe and D. Blei, “Supervised topic models,” in Advances in
neural information processing systems, 2008, pp. 121–128.
[8]
J. Zhu, A. Ahmed, and E. Xing, “Medlda: Maximum margin
supervised topic models,” J. Mach. Learn. Res., vol. 13, no. 1, pp.
2237–2278, 2012.
[9]
R. Snow, B. O’Connor, D. Jurafsky, and A. Ng, “Cheap and fast
- but is it good?: Evaluating non-expert annotations for natural
language tasks,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2008, pp. 254–263.
[10] F. Rodrigues, F. Pereira, and B. Ribeiro, “Learning from multiple
annotators: distinguishing good from random labelers,” Pattern
Recognition Letters, pp. 1428–1436, 2013.
[11] P. Welinder, S. Branson, P. Perona, and S. Belongie, “The multidi-
mensional wisdom of crowds,” in Advances in neural information
processing systems, 2010, pp. 2424–2432.
[12] Y. Yan, R. Rosales, G. Fung, R. Subramanian, and J. Dy, “Learning
from multiple annotators with varying expertise,” Mach. Learn.,
vol. 95, no. 3, pp. 291–327, 2014.
[13] F. Rodrigues, M. Lourenc¸o, F. Pereira, and B. Ribeiro, “Learning
supervised topic models from crowds,” in Proc. of the Third AAAI
Conf. on Human Computation and Crowdsourcing, 2015, pp. 160–168.
[14] M. Hoffman, D. Blei, C. Wang, and J. Paisley, “Stochastic varia-
tional inference,” J. Mach. Learn. Res., vol. 14, pp. 1303–1347, 2013.
[15] S. Lacoste-Julien, F. Sha, and M. Jordan, “Disclda: Discriminative
learning for dimensionality reduction and classification,” in Ad-
vances in neural information processing systems, 2009, pp. 897–904.
[16] D. Ramage, D. Hall, R. Nallapati, and C. Manning, “Labeled LDA:
A supervised topic model for credit attribution in multi-labeled
corpora,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2009, pp. 248–256.
[17] D. Mimno and A. McCallum, “Topic models conditioned on
arbitrary features with dirichlet-multinomial regression,” Proc. of
The 24th Conf. on Uncertainty in Artificial Intelligence, 2008.
[18] M. Rabinovich and D. Blei, “The inverse regression topic model,”
in Proc. of The 31st Int. Conf. on Machine Learning, 2014, pp. 199–207.
[19] M. Taddy, “Multinomial inverse regression for text analysis,” J. of
the American Statistical Assoc., vol. 108, no. 503, pp. 755–770, 2013.
[20] A. P. Dawid and A. M. Skene, “Maximum likelihood estimation
of observer error-rates using the EM algorithm,” J. of the Royal
Statistical Society. Series C, vol. 28, no. 1, pp. 20–28, 1979.
[21] P. Smyth, U. Fayyad, M. Burl, P. Perona, and P. Baldi, “Inferring
ground truth from subjective labelling of venus images,” in Ad-
vances in Neural Information Processing Systems, 1995, pp. 1085–1092.
[22] V. Raykar, S. Yu, L. Zhao, G. Valadez, C. Florin, L. Bogoni, and
L. Moy, “Learning from Crowds,” J. Mach. Learn. Res, pp. 1297–
1322, 2010.
[23] F. Rodrigues, F. Pereira, and B. Ribeiro, “Gaussian process classi-
fication and active learning with multiple annotators,” in Proc. of
the 31st Int. Conf. on Machine Learning, 2014, pp. 433–441.
[24] P. Groot, A. Birlutiu, and T. Heskes, “Learning from multiple
annotators with Gaussian processes,” in Proc. of the Int. Conf. on
Artificial Neural Networks, vol. 6792, 2011, pp. 159–164.
[25] Y. Yan, R. Rosales, G. Fung, M. Schmidt, G. Valadez, L. Bogoni,
L. Moy, and J. Dy, “Modeling annotator expertise: Learning when
everybody knows a bit of something,” J. Mach. Learn. Res, vol. 9,
pp. 932–939, 2010.
[26] M. I. Jordan, Z. Ghahramani, T. Jaakkola, and L. Saul, “An in-
troduction to variational methods for graphical models,” Mach.
Learn., vol. 37, no. 2, pp. 183–233, 1999.
[27] J. Chuang, S. Gupta, C. Manning, and J. Heer, “Topic model
diagnostics: Assessing domain relevance via topical alignment.”
in Proc. of the 30th Int. Conf. on Machine Learning, 2013, pp. 612–620.
[28] J. Nocedal and S. Wright, Numerical Optimization. World Sc., 2006.
[29] H. Robbins and S. Monro, “A stochastic approximation method,”
Ann. Math. Statist., vol. 22, no. 3, pp. 400–407, 09 1951.
[30] K. Lang, “Newsweeder: Learning to filter netnews,” in Proc. of the
Int. Conf. on Machine Learning, 1995, pp. 331–339.
[31] D. Lewis, “Reuters-21578 text categorization test collection. distri-
bution 1.0. readme file (version 1.2),” 1997. [Online]. Available:
www.daviddlewis.com/resources/testcollections/reuters21578/
[32] B. Russell, A. Torralba, K. Murphy, and W. Freeman, “Labelme:
a database and web-based tool for image annotation,” Int. J. of
Computer Vision, vol. 77, no. 1-3, pp. 157–173, 2008.",A diagram of a woman with long hair.,A detailed layout showing the eye and smile of a woman with red lipstick on her lips and shoulder length.,"A comprehensive technical explanation of the beauty of a woman with a red lip and long, straight hair with bangs on her head and shoulders, smiling at the camera in front of a restaurant table with boats in the background and windows.","[6]
E. Airoldi, D. Blei, S. Fienberg, and E. Xing, “Combining stochastic
block models and mixed membership for statistical network anal-
ysis,” in Statistical Network Analysis.
Springer, 2007, pp. 57–74.
[7]
J. Mcauliffe and D. Blei, “Supervised topic models,” in Advances in
neural information processing systems, 2008, pp. 121–128.
[8]
J. Zhu, A. Ahmed, and E. Xing, “Medlda: Maximum margin
supervised topic models,” J. Mach. Learn. Res., vol. 13, no. 1, pp.
2237–2278, 2012.
[9]
R. Snow, B. O’Connor, D. Jurafsky, and A. Ng, “Cheap and fast
- but is it good?: Evaluating non-expert annotations for natural
language tasks,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2008, pp. 254–263.
[10] F. Rodrigues, F. Pereira, and B. Ribeiro, “Learning from multiple
annotators: distinguishing good from random labelers,” Pattern
Recognition Letters, pp. 1428–1436, 2013.
[11] P. Welinder, S. Branson, P. Perona, and S. Belongie, “The multidi-
mensional wisdom of crowds,” in Advances in neural information
processing systems, 2010, pp. 2424–2432.
[12] Y. Yan, R. Rosales, G. Fung, R. Subramanian, and J. Dy, “Learning
from multiple annotators with varying expertise,” Mach. Learn.,
vol. 95, no. 3, pp. 291–327, 2014.
[13] F. Rodrigues, M. Lourenc¸o, F. Pereira, and B. Ribeiro, “Learning
supervised topic models from crowds,” in Proc. of the Third AAAI
Conf. on Human Computation and Crowdsourcing, 2015, pp. 160–168.
[14] M. Hoffman, D. Blei, C. Wang, and J. Paisley, “Stochastic varia-
tional inference,” J. Mach. Learn. Res., vol. 14, pp. 1303–1347, 2013.
[15] S. Lacoste-Julien, F. Sha, and M. Jordan, “Disclda: Discriminative
learning for dimensionality reduction and classification,” in Ad-
vances in neural information processing systems, 2009, pp. 897–904.
[16] D. Ramage, D. Hall, R. Nallapati, and C. Manning, “Labeled LDA:
A supervised topic model for credit attribution in multi-labeled
corpora,” in Proc. of the Conf. on Empirical Methods in Natural
Language Processing, 2009, pp. 248–256.
[17] D. Mimno and A. McCallum, “Topic models conditioned on
arbitrary features with dirichlet-multinomial regression,” Proc. of
The 24th Conf. on Uncertainty in Artificial Intelligence, 2008.
[18] M. Rabinovich and D. Blei, “The inverse regression topic model,”
in Proc. of The 31st Int. Conf. on Machine Learning, 2014, pp. 199–207.
[19] M. Taddy, “Multinomial inverse regression for text analysis,” J. of
the American Statistical Assoc., vol. 108, no. 503, pp. 755–770, 2013.
[20] A. P. Dawid and A. M. Skene, “Maximum likelihood estimation
of observer error-rates using the EM algorithm,” J. of the Royal
Statistical Society. Series C, vol. 28, no. 1, pp. 20–28, 1979.
[21] P. Smyth, U. Fayyad, M. Burl, P. Perona, and P. Baldi, “Inferring
ground truth from subjective labelling of venus images,” in Ad-
vances in Neural Information Processing Systems, 1995, pp. 1085–1092.
[22] V. Raykar, S. Yu, L. Zhao, G. Valadez, C. Florin, L. Bogoni, and
L. Moy, “Learning from Crowds,” J. Mach. Learn. Res, pp. 1297–
1322, 2010.
[23] F. Rodrigues, F. Pereira, and B. Ribeiro, “Gaussian process classi-
fication and active learning with multiple annotators,” in Proc. of
the 31st Int. Conf. on Machine Learning, 2014, pp. 433–441.
[24] P. Groot, A. Birlutiu, and T. Heskes, “Learning from multiple
annotators with Gaussian processes,” in Proc. of the Int. Conf. on
Artificial Neural Networks, vol. 6792, 2011, pp. 159–164.
[25] Y. Yan, R. Rosales, G. Fung, M. Schmidt, G. Valadez, L. Bogoni,
L. Moy, and J. Dy, “Modeling annotator expertise: Learning when
everybody knows a bit of something,” J. Mach. Learn. Res, vol. 9,
pp. 932–939, 2010.
[26] M. I. Jordan, Z. Ghahramani, T. Jaakkola, and L. Saul, “An in-
troduction to variational methods for graphical models,” Mach.
Learn., vol. 37, no. 2, pp. 183–233, 1999.
[27] J. Chuang, S. Gupta, C. Manning, and J. Heer, “Topic model
diagnostics: Assessing domain relevance via topical alignment.”
in Proc. of the 30th Int. Conf. on Machine Learning, 2013, pp. 612–620.
[28] J. Nocedal and S. Wright, Numerical Optimization. World Sc., 2006.
[29] H. Robbins and S. Monro, “A stochastic approximation method,”
Ann. Math. Statist., vol. 22, no. 3, pp. 400–407, 09 1951.
[30] K. Lang, “Newsweeder: Learning to filter netnews,” in Proc. of the
Int. Conf. on Machine Learning, 1995, pp. 331–339.
[31] D. Lewis, “Reuters-21578 text categorization test collection. distri-
bution 1.0. readme file (version 1.2),” 1997. [Online]. Available:
www.daviddlewis.com/resources/testcollections/reuters21578/
[32] B. Russell, A. Torralba, K. Murphy, and W. Freeman, “Labelme:
a database and web-based tool for image annotation,” Int. J. of
Computer Vision, vol. 77, no. 1-3, pp. 157–173, 2008.",0.725,0.2619,0.4934,825,1029,0.802,40f205daef6b5e08f816e1828f663f3f,images/2018/arxiv_0000438.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000439,Figure 439,scientific_figure,Learning Supervised Topic Models for Classification and Regression from Crowds,arXiv Research Authors,1808.05902v1,stat,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Learning Supervised Topic Models for Classification and Regression from Crowds. arXiv:1808.05902v1,"IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. X, NO. X, XXXX
15",A diagram of a man with a cell phone.,"A detailed layout showing the face of a man with a red shirt and earphones on his head that shows key features, attributes, and data points in.","A comprehensive technical explanation of the camera and its image of a man taking a selfie with his cell phone in front of him, and behind him is a picture of him is an advertisement for a sign that reads, there is a door.","IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. X, NO. X, XXXX
15",0.81,0.2709,0.5404,300,375,0.8,78816b124f25e0c3054f750679a56cfb,images/2018/arxiv_0000439.png,https://arxiv.org/pdf/1808.05902v1.pdf
arxiv_0000440,Figure 440,scientific_figure,Context-Dependent Diffusion Network for Visual Relationship Detection,arXiv Research Authors,1809.06213v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Context-Dependent Diffusion Network for Visual Relationship Detection. arXiv:1809.06213v1,"Figure 1: Motivation. The relationship triplets are often se-
mantically related to each other, e.g., person-eat-fruit, boy-
eat-sandwich, girl-eat-orange, people-eat-fruit, etc. The pro-
duced cliques of {person, boy, girl, · · · } and {sandwich, or-
ange, fruit, · · · } should have tight internal correlations,
which may be utilized to derive (new) visual relationships.
To this end, we build the possibility graph of object se-
mantics by using language priors from training data, and
then transmit/exchange information via edge connections,
which refers to diffusion in this paper. Note that these two
subgraphs only exhibit a few examples for better observa-
tion.",A diagram of a group of objects with.,"A detailed layout showing the different types of objects in the same network diagram, including a brain, eye, brain, and eye.","A comprehensive technical explanation of a networked system of objects and objects, including the brain, eye, and body of an object and the brain of the body of a person, from which is represented in the same area of the object.","Figure 1: Motivation. The relationship triplets are often se-
mantically related to each other, e.g., person-eat-fruit, boy-
eat-sandwich, girl-eat-orange, people-eat-fruit, etc. The pro-
duced cliques of {person, boy, girl, · · · } and {sandwich, or-
ange, fruit, · · · } should have tight internal correlations,
which may be utilized to derive (new) visual relationships.
To this end, we build the possibility graph of object se-
mantics by using language priors from training data, and
then transmit/exchange information via edge connections,
which refers to diffusion in this paper. Note that these two
subgraphs only exhibit a few examples for better observa-
tion.",0.87,0.2878,0.5789,566,360,1.572,ce485fb91f9d1b8bc3d6e523d641175f,images/2018/arxiv_0000440.png,https://arxiv.org/pdf/1809.06213v1.pdf
arxiv_0000441,Figure 441,scientific_figure,Context-Dependent Diffusion Network for Visual Relationship Detection,arXiv Research Authors,1809.06213v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Context-Dependent Diffusion Network for Visual Relationship Detection. arXiv:1809.06213v1,"Figure 2: The proposed CDDN architecture of visual relationship detection. We use two types of global context information:
semantic priors and spatial scenes. The corresponding graphs are built to capture internal correlations across object instances.
The diffusion is used to adaptively propagate context information such that the predicates can be well estimated. Details can
be found in Section 3. the training stage, the relationship triplets are known with pre-
annotated objects and relations. In testing, we first perform object
detection by some detectors (e.g., faster R-CNN [33]) to acquire
the locations, labels and confidence scores for all possible objects,
and then send them into this network to estimate the predicates
by ranking scores. classification problem. Furthermore, some variants of graph convo-
lution networks are also applied into many fields, such as skeleton-
based action recognition [18, 19], goodrecommendation system [42],
EEG-based emotion recognition [36], and so on. In contrast, in this
paper we most focus on the task of visual relationship detection,
and intorduce/revise graph convolution (concretely graph diffu-
sion) to cater to such a task. 3.2
Features 3
THE METHOD Visual appearance plays an important role in distinguishing the
categories of objects and understanding relationships. Let bs =
(xs;ys;ws ;hs) and bo = (xo;yo;wo;ho) denote the tuples of coor-
dinates, width and heights of the detected bounding boxes respec-
tively for a subject and an object. Considering a relative spatial
position, we also employ the union bounding box of both regions.
Instead of directly using the union, we perform the spatial mask
with regard to the subject and object on the union box, as used
in the method [23], such that the relative spatial layouts are pre-
served. For the subject, the spatial mask suppresses all positions
into zeros except for the subject region. Similarly, we conduct the
spatial mask for the object region. Thus we can obtain two masked
bounding boxes b′s, b′o on the union of the subject and object. Then
we employ convolution neural network as the backbone and use
RoI pooling to crop out the features of bs, bo, b′s, b′o. In our exper-
iments, we choose VGG16 [35] to extract visual features from the
last convolutional layer and then feed RoI features into two fully
connected layers.
Language prior is another important strategy to promote visual
relationship detection. If only using visual appearance features, the
estimation of predicates might be vague sometimes or very diffi-
cult due to the large diversity of relationships. Nevertheless, se-
mantic priors can alleviate this problem to some extend and mean-
time make the inference with a better generality. To use the visual
manifestation across different object categories, we embed each
object category into a vector space. Here we use word vector em-
bedding [27] as the off-the-shelf language model to acquire the In this section, we first overview the entire CDDN architecture,
and then introduce four submodules: feature extraction, object as-
sociation, diffusion layer and ranking loss. 3.1
Overview In visual relationship detection, we need to detect those possible
objects and determine the predicates of each pair of them. Let O
and P denote the object set and predicate set respectively, then the
relationship set can be defined as R = {(s,p,o)|s,o ∈O,p ∈P},
where s and o are respectively the subject and the object in a re-
lationship triplet. To better represent objects, we take two types
of features: visual appearance and semantic embedding. As shown
in Fig. 2, there are two stream pipelines. The top pipeline encodes
word semantic features to capture the semantic similarity in lan-
guage, while the bottom extracts expressive visual appearance fea-
tures through a convolutional neural network. The diffusion block
(Section 3.4) takes two inputs: i) graphs of semantic priors or spa-
tial scenes; and ii) object features of CNN output or word embed-
ding. Given a pair of objects, we can perform the diffusion opera-
tion on semantic/spatial scene graphs to integrate global context
information. For the construction of graphs, the detailed introduc-
tion can be found in Section 3.3. After diffusion, for a pair of objects
(s,o), we can obtain the diffused visual features and semantic fea-
tures for s,o. The two types of features are concatenated and then
fed into a structural ranking loss to decide the predicates. During",A diagram of a square with the words.,A detailed layout showing the two sections of a paper with the text spotti scene graph in the center.,"A comprehensive technical explanation of the spatial scene graph in photoshopped graphics and animation, including a series of diagrams and a text box with a picture inside it ' s shape ' s and text area ' s name ' s.","Figure 2: The proposed CDDN architecture of visual relationship detection. We use two types of global context information:
semantic priors and spatial scenes. The corresponding graphs are built to capture internal correlations across object instances.
The diffusion is used to adaptively propagate context information such that the predicates can be well estimated. Details can
be found in Section 3. the training stage, the relationship triplets are known with pre-
annotated objects and relations. In testing, we first perform object
detection by some detectors (e.g., faster R-CNN [33]) to acquire
the locations, labels and confidence scores for all possible objects,
and then send them into this network to estimate the predicates
by ranking scores. classification problem. Furthermore, some variants of graph convo-
lution networks are also applied into many fields, such as skeleton-
based action recognition [18, 19], goodrecommendation system [42],
EEG-based emotion recognition [36], and so on. In contrast, in this
paper we most focus on the task of visual relationship detection,
and intorduce/revise graph convolution (concretely graph diffu-
sion) to cater to such a task. 3.2
Features 3
THE METHOD Visual appearance plays an important role in distinguishing the
categories of objects and understanding relationships. Let bs =
(xs;ys;ws ;hs) and bo = (xo;yo;wo;ho) denote the tuples of coor-
dinates, width and heights of the detected bounding boxes respec-
tively for a subject and an object. Considering a relative spatial
position, we also employ the union bounding box of both regions.
Instead of directly using the union, we perform the spatial mask
with regard to the subject and object on the union box, as used
in the method [23], such that the relative spatial layouts are pre-
served. For the subject, the spatial mask suppresses all positions
into zeros except for the subject region. Similarly, we conduct the
spatial mask for the object region. Thus we can obtain two masked
bounding boxes b′s, b′o on the union of the subject and object. Then
we employ convolution neural network as the backbone and use
RoI pooling to crop out the features of bs, bo, b′s, b′o. In our exper-
iments, we choose VGG16 [35] to extract visual features from the
last convolutional layer and then feed RoI features into two fully
connected layers.
Language prior is another important strategy to promote visual
relationship detection. If only using visual appearance features, the
estimation of predicates might be vague sometimes or very diffi-
cult due to the large diversity of relationships. Nevertheless, se-
mantic priors can alleviate this problem to some extend and mean-
time make the inference with a better generality. To use the visual
manifestation across different object categories, we embed each
object category into a vector space. Here we use word vector em-
bedding [27] as the off-the-shelf language model to acquire the In this section, we first overview the entire CDDN architecture,
and then introduce four submodules: feature extraction, object as-
sociation, diffusion layer and ranking loss. 3.1
Overview In visual relationship detection, we need to detect those possible
objects and determine the predicates of each pair of them. Let O
and P denote the object set and predicate set respectively, then the
relationship set can be defined as R = {(s,p,o)|s,o ∈O,p ∈P},
where s and o are respectively the subject and the object in a re-
lationship triplet. To better represent objects, we take two types
of features: visual appearance and semantic embedding. As shown
in Fig. 2, there are two stream pipelines. The top pipeline encodes
word semantic features to capture the semantic similarity in lan-
guage, while the bottom extracts expressive visual appearance fea-
tures through a convolutional neural network. The diffusion block
(Section 3.4) takes two inputs: i) graphs of semantic priors or spa-
tial scenes; and ii) object features of CNN output or word embed-
ding. Given a pair of objects, we can perform the diffusion opera-
tion on semantic/spatial scene graphs to integrate global context
information. For the construction of graphs, the detailed introduc-
tion can be found in Section 3.3. After diffusion, for a pair of objects
(s,o), we can obtain the diffused visual features and semantic fea-
tures for s,o. The two types of features are concatenated and then
fed into a structural ranking loss to decide the predicates. During",0.7707,0.3153,0.543,957,682,1.403,9e3f455fa84c4f8ccbfea1c58081f07a,images/2018/arxiv_0000441.png,https://arxiv.org/pdf/1809.06213v1.pdf
arxiv_0000442,Figure 442,scientific_figure,Context-Dependent Diffusion Network for Visual Relationship Detection,arXiv Research Authors,1809.06213v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Context-Dependent Diffusion Network for Visual Relationship Detection. arXiv:1809.06213v1,NULL,A diagram of a square with a square.,"A detailed layout showing the shape of a square with a blue border and a small rounded corner, with the words spatial scene graph.","A comprehensive technical explanation of the typical scene graph for a film production company, from the first and second half of the 20th century to the present day of the year of the century, as well as soon as the present,.",NULL,0.81,0.3141,0.5621,319,230,1.387,88ad5c247d7938e1eddfb4be66e4bf8c,images/2018/arxiv_0000442.png,https://arxiv.org/pdf/1809.06213v1.pdf
arxiv_0000443,Figure 443,scientific_figure,Context-Dependent Diffusion Network for Visual Relationship Detection,arXiv Research Authors,1809.06213v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Context-Dependent Diffusion Network for Visual Relationship Detection. arXiv:1809.06213v1,NULL,A diagram of the word embedding and.,"A detailed layout showing the word embedding and the image of a blank frame with text and a shadow that shows key features, attributes, and data.","A comprehensive technical explanation of semi - texting in the classroom of graphic graphics, inc, with examples of using the basic formatting tool to create a word embedding for each subject or image for a text, with a picture.",NULL,0.87,0.307,0.5885,957,682,1.403,ac5c1b9ace75861bd2a52872a772809b,images/2018/arxiv_0000443.png,https://arxiv.org/pdf/1809.06213v1.pdf
arxiv_0000444,Figure 444,scientific_figure,Context-Dependent Diffusion Network for Visual Relationship Detection,arXiv Research Authors,1809.06213v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Context-Dependent Diffusion Network for Visual Relationship Detection. arXiv:1809.06213v1,NULL,A diagram of the same color as shown.,"A detailed layout showing the layout of a website page with a white background and blue font, including a white.","A comprehensive technical explanation of the standard waterproofing system for the san antonio, california area, including the use of the new paint and flooring method to maintain the original waterproofer ' s surface to create the water.",NULL,0.75,0.2947,0.5223,319,229,1.393,9021e0ef1a9f1ca1b98231382e702da7,images/2018/arxiv_0000444.jpeg,https://arxiv.org/pdf/1809.06213v1.pdf
arxiv_0000445,Figure 445,scientific_figure,Context-Dependent Diffusion Network for Visual Relationship Detection,arXiv Research Authors,1809.06213v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Context-Dependent Diffusion Network for Visual Relationship Detection. arXiv:1809.06213v1,"Figure 2: The proposed CDDN architecture of visual relationship detection. We use two types of global context information:
semantic priors and spatial scenes. The corresponding graphs are built to capture internal correlations across object instances.
The diffusion is used to adaptively propagate context information such that the predicates can be well estimated. Details can
be found in Section 3.",A diagram of a group of people riding.,"A detailed layout showing the different stages of a horse riding activity, including a harness, a training course, and two other activities.","A comprehensive technical explanation of the basic design of the camera system for video cameras and films, including the image of the film camera and the image on screen, the picture, from the left side, is shown in the right.","Figure 2: The proposed CDDN architecture of visual relationship detection. We use two types of global context information:
semantic priors and spatial scenes. The corresponding graphs are built to capture internal correlations across object instances.
The diffusion is used to adaptively propagate context information such that the predicates can be well estimated. Details can
be found in Section 3.",0.81,0.182,0.496,985,343,2.872,ef7f991e0fd01d980170bfe2eff15d58,images/2018/arxiv_0000445.png,https://arxiv.org/pdf/1809.06213v1.pdf
arxiv_0000446,Figure 446,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 1. Example of MTLE-based video captioning. Ground truth
captions are compared with a semantic distance function (SDF) to
find semantically disjoint caption samples, which are passed to our
MTLE method to produce a a video’s caption. where typical encoder-decoder frameworks are employed. Learning and encoding visual feature representations for
video analysis and specifically video captioning is challeng-
ing. Some of these challenges come from the complexity
and nature of the data where video frames displayed as im-
ages in a time sequence add a temporal dimension to a much
larger challenge of recognizing and detecting objects in a
per frame basis. Because of the variable length of its tem- 1",A diagram of a woman with a cell phone.,A detailed layout showing three different shots of a woman and man holding papers and talking to each other.,"A comprehensive technical explanation of the camera ' s effects in a film strip, and a film reel image of a woman at a desk with a camera attached to her laptop computer and a man at her lap top, with a plant.","Figure 1. Example of MTLE-based video captioning. Ground truth
captions are compared with a semantic distance function (SDF) to
find semantically disjoint caption samples, which are passed to our
MTLE method to produce a a video’s caption. where typical encoder-decoder frameworks are employed. Learning and encoding visual feature representations for
video analysis and specifically video captioning is challeng-
ing. Some of these challenges come from the complexity
and nature of the data where video frames displayed as im-
ages in a time sequence add a temporal dimension to a much
larger challenge of recognizing and detecting objects in a
per frame basis. Because of the variable length of its tem- 1",0.7357,0.3146,0.5252,1110,269,4.126,80fbaadaf872e10050b89bcf34964077,images/2018/arxiv_0000446.jpeg,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000447,Figure 447,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 1. Example of MTLE-based video captioning. Ground truth
captions are compared with a semantic distance function (SDF) to
find semantically disjoint caption samples, which are passed to our
MTLE method to produce a a video’s caption.",A diagram of how to use the phone in.,"A detailed layout showing the steps of an effective meeting in a phone call center, including a phone.",A comprehensive technical explanation of a callout for a woman and a man speaking in an office a man is meeting on an office b a woman speaks on the phone c a woman on the telephone c woman speaks in the phone.,"Figure 1. Example of MTLE-based video captioning. Ground truth
captions are compared with a semantic distance function (SDF) to
find semantically disjoint caption samples, which are passed to our
MTLE method to produce a a video’s caption.",0.7143,0.3243,0.5193,476,404,1.178,bcfef50f08de34ec03e62a04a38112c3,images/2018/arxiv_0000447.png,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000448,Figure 448,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 2. Overview of our MTLE method. Video frames are passed to a CNN encoder for feature vector generation. From the encoder,
feature vectors are passed to an RNN-based bi-directional attention encoder. The output of this process is the concatenated visual feature en-
coding of the video, which is passed to a multitask conditional decoder with soft attention. The parameters of the RNN-based bi-directional
attention encoder are denoted by θe, and are trained simultaneous to the multitask decoder. Green arrows denote back-propagation.",A diagram of the process for the.,"A detailed layout showing the phases of a cellular transport system, including the initial and initial stages of the cell.","A comprehensive technical explanation of the three - stage process for the first step in the process, including the initial and initial steps of each process, as described by the two - step diagram, and the second step, the following.","Figure 2. Overview of our MTLE method. Video frames are passed to a CNN encoder for feature vector generation. From the encoder,
feature vectors are passed to an RNN-based bi-directional attention encoder. The output of this process is the concatenated visual feature en-
coding of the video, which is passed to a multitask conditional decoder with soft attention. The parameters of the RNN-based bi-directional
attention encoder are denoted by θe, and are trained simultaneous to the multitask decoder. Green arrows denote back-propagation.",0.75,0.2489,0.4995,789,285,2.768,0c1845f846dbb99a148843259d8c7524,images/2018/arxiv_0000448.png,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000449,Figure 449,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 3. Semantic space representation with t-SNE embedding of
skip-thought vectors of captions in different videos. Each dot in
the graph represents a caption of a video represented by a particu-
lar color.",A diagram of the semantic space.,A detailed layout showing the different types of semantic space clusters and the different sizes of them in the image.,"A comprehensive technical explanation of semantic space clusters, with examples of the data and visual elements in a single file, including the data visual file, and the data file and the image in each one file, in the file,.","Figure 3. Semantic space representation with t-SNE embedding of
skip-thought vectors of captions in different videos. Each dot in
the graph represents a caption of a video represented by a particu-
lar color.",0.75,0.3649,0.5575,402,329,1.222,21a50fb47da8cc97eb6f0ca04c296511,images/2018/arxiv_0000449.png,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000450,Figure 450,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 4. Human judgment histograms of subjective scoring. His-
tograms indicate the percentage of errors classified as minor or
major. Notice our method contains the smaller percentage of ma-
jor and minor errors. 4.3.1
Qualitative Results",A diagram of mixing ingredients in a.,"A detailed layout showing the process of mixing and stirring ingredients in a blender, with a hand holding a whisk.",A comprehensive technical explanation of mixing ingredients in a blender with a large spoon and a small bowl for making something that looks like a smoothie or dessert is ready to be made from scratchy? with a cake mix -.,"Figure 4. Human judgment histograms of subjective scoring. His-
tograms indicate the percentage of errors classified as minor or
major. Notice our method contains the smaller percentage of ma-
jor and minor errors. 4.3.1
Qualitative Results",0.75,0.3122,0.5311,1547,250,6.188,a4d29fb6703fdce19f7218d85df72f69,images/2018/arxiv_0000450.jpeg,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000451,Figure 451,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 5 shows qualitative results of our approach for
uni/multi-label videos on the MSVD, MSR-VTT, LSMDC
and TRECVID datasets. It is worth to notice that in some in-
stances our system performs near human performance such
as the second video of TRECVID 5(d) and the first video
of LSMDC 5(c). In the second video of MSR-VTT 5(b)
our method even helps recognize the gender of the person,
something the ground truth did not provide. Figure 5 shows
some of our best results. We include in the supplementary
material some fail cases and other comparisons and details
that could not be included in this draft because of space con-
straints.
Source Code: The code use for the competition was re-
leased at: https://github.com/OSUPCVLab/
VideoToTextDNN 5. Conclusion In this paper we have presented a novel multitask
encoder-decoder framework for semantic video and movie
description. Our method helps improve a video feature en-
coder by leveraging the diversity of captions and a multitask
framework to solve a multitask loss function through a con-
vex optimization. Our method shows promising results and
in a human evaluation was ranked the highest and most use-
ful among other methods for helping the visually impaired. 6. Acknowledgement We would like to thank Li Yao for his code and support
with such. Thanks to Anna Rohrbach and the organizers of
the LSMDC 2017 competition for providing us with data re-
sults and figures from the competition for the publication of
this paper. We would also like to thank insightful comments
and discussions from Steven Rogers and Vincent Velten.
This work was funded by the ASEE SMART program.
The documentation has been approved for public release by",A diagram of a man in a suit and tie.,A detailed layout showing the same image of a person with colorful balloons on their head and a picture of a man in a suit.,"A comprehensive technical explanation of the different hats worn by the clowns in a circus show, including one with a hat and one with an umbrella, are shown in multiple images of clowns wearing a suit with a mask,.","Figure 5 shows qualitative results of our approach for
uni/multi-label videos on the MSVD, MSR-VTT, LSMDC
and TRECVID datasets. It is worth to notice that in some in-
stances our system performs near human performance such
as the second video of TRECVID 5(d) and the first video
of LSMDC 5(c). In the second video of MSR-VTT 5(b)
our method even helps recognize the gender of the person,
something the ground truth did not provide. Figure 5 shows
some of our best results. We include in the supplementary
material some fail cases and other comparisons and details
that could not be included in this draft because of space con-
straints.
Source Code: The code use for the competition was re-
leased at: https://github.com/OSUPCVLab/
VideoToTextDNN 5. Conclusion In this paper we have presented a novel multitask
encoder-decoder framework for semantic video and movie
description. Our method helps improve a video feature en-
coder by leveraging the diversity of captions and a multitask
framework to solve a multitask loss function through a con-
vex optimization. Our method shows promising results and
in a human evaluation was ranked the highest and most use-
ful among other methods for helping the visually impaired. 6. Acknowledgement We would like to thank Li Yao for his code and support
with such. Thanks to Anna Rohrbach and the organizers of
the LSMDC 2017 competition for providing us with data re-
sults and figures from the competition for the publication of
this paper. We would also like to thank insightful comments
and discussions from Steven Rogers and Vincent Velten.
This work was funded by the ASEE SMART program.
The documentation has been approved for public release by",0.75,0.2204,0.4852,1536,221,6.95,d4455b5d7692361a87a035ec12a8ddd5,images/2018/arxiv_0000451.jpeg,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000452,Figure 452,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 4. Human judgment histograms of subjective scoring. His-
tograms indicate the percentage of errors classified as minor or
major. Notice our method contains the smaller percentage of ma-
jor and minor errors.",A diagram of the percentage of people.,"A detailed layout showing the various types of human attachments in the body of a person, including.","A comprehensive technical explanation of the internet and the internet in the uk, by language and language, as well as other languages and languages, for each country, including english and english and french, from the united kingdom of the united.","Figure 4. Human judgment histograms of subjective scoring. His-
tograms indicate the percentage of errors classified as minor or
major. Notice our method contains the smaller percentage of ma-
jor and minor errors.",0.7071,0.2212,0.4641,468,234,2.0,5bb9159023d43ae99de4196901dbf9e7,images/2018/arxiv_0000452.png,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000453,Figure 453,scientific_figure,MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description,arXiv Research Authors,1809.07257v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). MTLE: A Multitask Learning Encoder of Visual Feature Representations for Video and Movie Description. arXiv:1809.07257v1,"Figure 5. Qualitative results of our method on 5(a) MSVD, 5(b)
MSR-VTT, 5(c) LSMDC, and 5(d) TRECVID datasets.",A diagram of the various types of.,"A detailed layout showing the different views of a video game screen, including a camera and a player.","A comprehensive technical explanation of the various types of cameras used to film a movie set, including a camera and a screen shot of a man in a suit and woman in a white dress with a black coat and a red tie.","Figure 5. Qualitative results of our method on 5(a) MSVD, 5(b)
MSR-VTT, 5(c) LSMDC, and 5(d) TRECVID datasets.",0.7743,0.2932,0.5337,505,1080,0.468,cb72b182f31f35a3ddf6f5d4303540c0,images/2018/arxiv_0000453.png,https://arxiv.org/pdf/1809.07257v1.pdf
arxiv_0000454,Figure 454,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 1. Overview of the MMVR model. It consists of two
pre-trained modules – an image generator (G) that inputs a
latent representation ht and generates an image ˆx; and an im-
age captioner that inputs an image ˆx and generates a caption
ˆy. To update the latent vector ht, cross-entropy between the
generated caption ˆy and a ground truth caption y is used while
the weights for the generator and CNN are fixed. The forward pass is initiated by passing a random latent
vector ht into the image generator which generates an image
ˆx. The image captioner uses the generated image to create
a caption. Word-level cross entropy is used to determine the
error between the generated caption, ˆy and a ground truth cap-
tion y. This error is used to iteratively update ht (and thus ˆx),
while keeping all other components fixed. With each itera-
tion, ˆy approaches y, and the generated image ˆx serves as a
proxy for the target caption. The gradient associated with the
cross-entropy error is specified in (1).",A diagram of a forward propagation in.,A detailed layout showing the different components of a program and how they are used to perform it in a motion.,"A comprehensive technical explanation of the forward propagation and forward - vectoring of a vehicle ' s engine in an auto - vehicle carrier system, including a vw and a typical vehicle ' front - side - by - side ' s - view.","Fig. 1. Overview of the MMVR model. It consists of two
pre-trained modules – an image generator (G) that inputs a
latent representation ht and generates an image ˆx; and an im-
age captioner that inputs an image ˆx and generates a caption
ˆy. To update the latent vector ht, cross-entropy between the
generated caption ˆy and a ground truth caption y is used while
the weights for the generator and CNN are fixed. The forward pass is initiated by passing a random latent
vector ht into the image generator which generates an image
ˆx. The image captioner uses the generated image to create
a caption. Word-level cross entropy is used to determine the
error between the generated caption, ˆy and a ground truth cap-
tion y. This error is used to iteratively update ht (and thus ˆx),
while keeping all other components fixed. With each itera-
tion, ˆy approaches y, and the generated image ˆx serves as a
proxy for the target caption. The gradient associated with the
cross-entropy error is specified in (1).",0.75,0.29,0.52,861,765,1.125,10b4476c72ddbd82a87997e0102c6c92,images/2018/arxiv_0000454.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000455,Figure 455,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 1. Overview of the MMVR model. It consists of two
pre-trained modules – an image generator (G) that inputs a
latent representation ht and generates an image ˆx; and an im-
age captioner that inputs an image ˆx and generates a caption
ˆy. To update the latent vector ht, cross-entropy between the
generated caption ˆy and a ground truth caption y is used while
the weights for the generator and CNN are fixed.",A diagram of the three different types.,A detailed layout showing the basic components of a digital circuit and its corresponding functions for the system and the current.,"A comprehensive technical explanation of the multi - dimensional vector representation for the multimode network and the multileved network in a single source system, two - way circuit, multi - level system, three - dimensional system,.","Fig. 1. Overview of the MMVR model. It consists of two
pre-trained modules – an image generator (G) that inputs a
latent representation ht and generates an image ˆx; and an im-
age captioner that inputs an image ˆx and generates a caption
ˆy. To update the latent vector ht, cross-entropy between the
generated caption ˆy and a ground truth caption y is used while
the weights for the generator and CNN are fixed.",0.75,0.1978,0.4739,569,1014,0.561,28623ded8e06b187b84f04fbc40dcfbb,images/2018/arxiv_0000455.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000456,Figure 456,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 2. Examples of the YOLO object detection on generated
images. The bounding boxes and corresponding labels are
detections with confidence greater than 0.5 threshold. Human Evaluations –
We conduct human evaluations to
validate image generation. We collected 50 image-caption
pairs and asked 80 humans (not including any of the authors)",A diagram of a bird and a toilet with.,"A detailed layout showing the various types of objects that are being viewed in a picture frame, including an object with a blue arrow.","A comprehensive technical explanation of the camera ' s image of an animal and its surroundings, including an ostrich and a toilet, a large bird and a field of grass with a white bag and a bird on it ' s head.","Fig. 2. Examples of the YOLO object detection on generated
images. The bounding boxes and corresponding labels are
detections with confidence greater than 0.5 threshold. Human Evaluations –
We conduct human evaluations to
validate image generation. We collected 50 image-caption
pairs and asked 80 humans (not including any of the authors)",0.75,0.259,0.5045,1258,300,4.193,f3de13cb0e7ff498c8b48417903bc30b,images/2018/arxiv_0000456.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000457,Figure 457,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,Fig. 3. Examples of text-to-visual transformation.,A diagram of a bird perched on a.,A detailed layout showing four different photographs of a vase with flowers in it and a bird sitting on a window sill.,"A comprehensive technical explanation of painting and painting with photos of birds, flowers, and vases on the table and side of a window in front of it, and a window, with a photograph of a bird sitting on a wooden bench.",Fig. 3. Examples of text-to-visual transformation.,0.81,0.3334,0.5717,847,396,2.139,3697ec6dc9abb2689802bde759d08f98,images/2018/arxiv_0000457.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000458,Figure 458,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,Fig. 3. Examples of text-to-visual transformation.,A diagram of a cat sitting on top of a.,"A detailed layout showing the various aspects of the object, including the image and text, and the design.","A comprehensive technical explanation of the art of painting and drawing using lightrooms and shadowboxes, with a detailed visual approach and a step - by - by step - based approach to draw - by, step guide.",Fig. 3. Examples of text-to-visual transformation.,0.7286,0.2612,0.4949,478,989,0.483,189389f80cc1c3b8765c65372c1bc09d,images/2018/arxiv_0000458.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000459,Figure 459,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 2. Examples of the YOLO object detection on generated
images. The bounding boxes and corresponding labels are
detections with confidence greater than 0.5 threshold.",A diagram of the process of making a.,A detailed layout showing the steps of creating a model for a modern home kitchen with a large scale.,"A comprehensive technical explanation of the product and its uses in 3d printing, including a full size model of a ball, and a detailed design by a designer, and discussion on - page from a different perspective to the same perspective.","Fig. 2. Examples of the YOLO object detection on generated
images. The bounding boxes and corresponding labels are
detections with confidence greater than 0.5 threshold.",0.7707,0.1675,0.4691,569,1280,0.445,b068222bcd9880508041c2843815bb55,images/2018/arxiv_0000459.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000460,Figure 460,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 4. Examples of the text-to-image generation as condi-
tioned on varying number of input captions. We observe more
detailed images being synthesized with increase in captions.",A diagram of a picture of various food.,"A detailed layout showing a variety of pictures of different food items and beverages, including a bus, a bottle of wine, an orange juice, a plate.","A comprehensive technical explanation of a food and beverage photography course with photos by tim mullickford, photographer and author of the new york times magazine, december 2014, new york, ny, newyog, ny -.","Fig. 4. Examples of the text-to-image generation as condi-
tioned on varying number of input captions. We observe more
detailed images being synthesized with increase in captions.",0.75,0.2936,0.5218,784,668,1.174,f2adf4d36a600a252425dda54ccf53cb,images/2018/arxiv_0000460.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000461,Figure 461,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 5. Examples comparing the text-to-image for PPGN and
the BLEU-1 scaled cross-entropy. Even though slight im-
provements could be observed with the n-gram scaling, judg-
ing the image quality visually is very challenging.",A diagram of four pictures with.,"A detailed layout showing four different images of boats and flowers in water, including a cat in the boat, and a dog in a boat.","A comprehensive technical explanation of how to clean a boat in the water, and what it can do with it for you to get out of the water or on the boat, and not in the river, and use it representing all details, specifications, and configurations of.","Fig. 5. Examples comparing the text-to-image for PPGN and
the BLEU-1 scaled cross-entropy. Even though slight im-
provements could be observed with the n-gram scaling, judg-
ing the image quality visually is very challenging.",0.75,0.3454,0.5477,708,547,1.294,bfe71ebab50f3cf3182df29cdef0fc4d,images/2018/arxiv_0000461.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000462,Figure 462,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 5. Examples comparing the text-to-image for PPGN and
the BLEU-1 scaled cross-entropy. Even though slight im-
provements could be observed with the n-gram scaling, judg-
ing the image quality visually is very challenging.",A diagram of how to use an inflatable.,"A detailed layout showing the stages of a boat being used in the process of painting it, including flowers and plants.","A comprehensive technical explanation of the process of painting a boat on a beach near some water, two vessels that are both white and one has green leaves painted on, one is blue and one is red on, and one.","Fig. 5. Examples comparing the text-to-image for PPGN and
the BLEU-1 scaled cross-entropy. Even though slight im-
provements could be observed with the n-gram scaling, judg-
ing the image quality visually is very challenging.",0.75,0.3248,0.5374,568,465,1.222,1012c87a639e18b031ca0f36a35f1c75,images/2018/arxiv_0000462.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000463,Figure 463,scientific_figure,Semantically Invariant Text-to-Image Generation,arXiv Research Authors,1809.10274v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Semantically Invariant Text-to-Image Generation. arXiv:1809.10274v1,"Fig. 4. Examples of the text-to-image generation as condi-
tioned on varying number of input captions. We observe more
detailed images being synthesized with increase in captions.",A diagram of the functions of plants.,A detailed layout showing the table and its contents for each class of students to learn in their classroom.,"A comprehensive technical explanation of the general inequalty using the table 1 description and human actions on the text, 2, 4, 5, 6, 6 and 6, 7, 9, 7 and 8, 9 representing all details, specifications, and configurations of the system components.","Fig. 4. Examples of the text-to-image generation as condi-
tioned on varying number of input captions. We observe more
detailed images being synthesized with increase in captions.",0.7957,0.2784,0.5371,569,901,0.632,5096b6b78c498e6f8e305db28ad4738d,images/2018/arxiv_0000463.png,https://arxiv.org/pdf/1809.10274v1.pdf
arxiv_0000464,Figure 464,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 1. VQA with out-of-vocabulary answers. Given a set of labels in visual dataset A and a set of answers in VQA training set B,
we evaluate a model on VQA test set with answers a ∈A −B. External visual dataset provides a set of bounding box labels and visual
descriptions for all answers in VQA training set and test set. See text for details. task annotations, we propose an unsupervised task discov-
ery technique based on linguistic knowledge sources such
as structured lexical databases, e.g., WordNet [10] and re-
gion descriptions [22]. We claim that the proposed transfer
learning framework helps generalization in VQA with out-
of-vocabulary answers. The main contribution of our paper
is three-fold: question types appearing in the training set. This split is
constructed by repurposing an existing VQA dataset [2] or
by constructing a synthetic dataset [18]. The problem set-
ting studied in this paper is similar to [19, 36] in the sense
that out-of-vocabulary answers are used for testing, but un-
like the prior work, we formulate the problem as a transfer
learning, where out-of-vocabulary answers are learned from
external visual data.
External data are often employed in VQA for better gen-
eralization.
Convolutional neural networks [15, 23] pre-
trained on ImageNet [9] is a widely accepted standard for
diverse VQA models [12, 43]. As an alternative, object de-
tector [34] trained on the Visual Genome dataset [22] is em-
ployed to extract pretrained visual features [3]. Pretrained
language models such as word embeddings [33] or sentence
embeddings [21] are frequently used to initialize parameters
of question encoders [12, 30, 35]. Exploiting information
retrieval from knowledge base [6, 7] or external vision algo-
rithms [40] to provide additional inputs to VQA models are
investigated in [39, 40, 41]. Transfer between VQA datasets
is studied in [16]. Sharing aligned image-word representa-
tions between VQA models and image classifiers has been
proposed in [14] to exploit external visual data.
Transfer learning from external data to cope with out-
of-vocabulary words has hardly been studied in VQA, but
is actively investigated in novel object captioning [4, 28,
37, 44]. For example, [4] and [37] decompose image cap-
tioning task into visual classification and language mod-
eling, and exploit unpaired visual and linguistic data as
additional resources to train visual classifier and language
model, respectively. Recent approaches incorporate pointer
networks [38] and learn to point an index of word candi-
dates [44] or an associated region [28], where the word
candidates are detected by a multi-label classifier [44] or
an object detector [28] trained with external visual data.
However, these algorithms are not directly applicable to • We present a novel transfer learning algorithm for vi-
sual question answering based on a task conditional
visual classifier. • We propose an unsupervised task discovery technique
for learning task conditional visual classifiers without
explicit task annotations. • We show that the proposed approach handle out-of-
vocabulary answers through knowledge transfer from
visual dataset without question annotations. The rest of the paper is organized as follows. Section 2
discusses prior works related to our approach. We describe
the overall transfer learning framework in Section 3. Learn-
ing visual concepts by unsupervised task discovery is de-
scribed in Section 4. Section 5 analyzes experimental re-
sults and Section 6 makes our conclusion. 2. Related Works The standard VQA evaluation assumes identically dis-
tributed train and test set [5, 29, 45]. As this evaluation
setting turns out to be vulnerable to models exploiting bi-
ases in training set [13], several alternatives have been pro-
posed. One approach is to reduce observed biases either by
balancing answers for individual questions [13] or by pro-
viding different biases to train and test sets intentionally [1].
Another approach is to construct compositional generaliza-
tion split [2, 18] whose question and answer pairs in test set
are formed by novel compositions of visual concepts and",A diagram of a group of photos with.,A detailed layout showing the basic steps of taking a virtual tour in china and the following steps that can be taken.,"A comprehensive technical explanation of the vlog web site for students to use on their own computer, including a video guide and a training manual for teachers to learn it all kinds of videos and how to do it???.","Figure 1. VQA with out-of-vocabulary answers. Given a set of labels in visual dataset A and a set of answers in VQA training set B,
we evaluate a model on VQA test set with answers a ∈A −B. External visual dataset provides a set of bounding box labels and visual
descriptions for all answers in VQA training set and test set. See text for details. task annotations, we propose an unsupervised task discov-
ery technique based on linguistic knowledge sources such
as structured lexical databases, e.g., WordNet [10] and re-
gion descriptions [22]. We claim that the proposed transfer
learning framework helps generalization in VQA with out-
of-vocabulary answers. The main contribution of our paper
is three-fold: question types appearing in the training set. This split is
constructed by repurposing an existing VQA dataset [2] or
by constructing a synthetic dataset [18]. The problem set-
ting studied in this paper is similar to [19, 36] in the sense
that out-of-vocabulary answers are used for testing, but un-
like the prior work, we formulate the problem as a transfer
learning, where out-of-vocabulary answers are learned from
external visual data.
External data are often employed in VQA for better gen-
eralization.
Convolutional neural networks [15, 23] pre-
trained on ImageNet [9] is a widely accepted standard for
diverse VQA models [12, 43]. As an alternative, object de-
tector [34] trained on the Visual Genome dataset [22] is em-
ployed to extract pretrained visual features [3]. Pretrained
language models such as word embeddings [33] or sentence
embeddings [21] are frequently used to initialize parameters
of question encoders [12, 30, 35]. Exploiting information
retrieval from knowledge base [6, 7] or external vision algo-
rithms [40] to provide additional inputs to VQA models are
investigated in [39, 40, 41]. Transfer between VQA datasets
is studied in [16]. Sharing aligned image-word representa-
tions between VQA models and image classifiers has been
proposed in [14] to exploit external visual data.
Transfer learning from external data to cope with out-
of-vocabulary words has hardly been studied in VQA, but
is actively investigated in novel object captioning [4, 28,
37, 44]. For example, [4] and [37] decompose image cap-
tioning task into visual classification and language mod-
eling, and exploit unpaired visual and linguistic data as
additional resources to train visual classifier and language
model, respectively. Recent approaches incorporate pointer
networks [38] and learn to point an index of word candi-
dates [44] or an associated region [28], where the word
candidates are detected by a multi-label classifier [44] or
an object detector [28] trained with external visual data.
However, these algorithms are not directly applicable to • We present a novel transfer learning algorithm for vi-
sual question answering based on a task conditional
visual classifier. • We propose an unsupervised task discovery technique
for learning task conditional visual classifiers without
explicit task annotations. • We show that the proposed approach handle out-of-
vocabulary answers through knowledge transfer from
visual dataset without question annotations. The rest of the paper is organized as follows. Section 2
discusses prior works related to our approach. We describe
the overall transfer learning framework in Section 3. Learn-
ing visual concepts by unsupervised task discovery is de-
scribed in Section 4. Section 5 analyzes experimental re-
sults and Section 6 makes our conclusion. 2. Related Works The standard VQA evaluation assumes identically dis-
tributed train and test set [5, 29, 45]. As this evaluation
setting turns out to be vulnerable to models exploiting bi-
ases in training set [13], several alternatives have been pro-
posed. One approach is to reduce observed biases either by
balancing answers for individual questions [13] or by pro-
viding different biases to train and test sets intentionally [1].
Another approach is to construct compositional generaliza-
tion split [2, 18] whose question and answer pairs in test set
are formed by novel compositions of visual concepts and",0.75,0.2566,0.5033,2598,870,2.986,62b221c9cb39f9f90522946002f05e65,images/2018/arxiv_0000464.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000465,Figure 465,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 1. VQA with out-of-vocabulary answers. Given a set of labels in visual dataset A and a set of answers in VQA training set B,
we evaluate a model on VQA test set with answers a ∈A −B. External visual dataset provides a set of bounding box labels and visual
descriptions for all answers in VQA training set and test set. See text for details.",A diagram of a number of different.,"A detailed layout showing the main features of a web page for a video game console, with a menu highlighted.","A comprehensive technical explanation of a website design for the internet market and its competitors, including a number of images and texting options, including the title and description tabs on the page below the image, above the image.","Figure 1. VQA with out-of-vocabulary answers. Given a set of labels in visual dataset A and a set of answers in VQA training set B,
we evaluate a model on VQA test set with answers a ∈A −B. External visual dataset provides a set of bounding box labels and visual
descriptions for all answers in VQA training set and test set. See text for details.",0.7957,0.2855,0.5406,1224,476,2.571,b014dd59172ccbe85394278cec69f2e5,images/2018/arxiv_0000465.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000466,Figure 466,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 2. Overview of the proposed algorithm. (a) Unsupervised task discovery samples a task specification for a sampled visual data
(a, I, b), where I, b and a are an image, a bounding box and an label (answer), respectively. It leverages linguistic knowledge sources
such as visual description d and WordNet. (b) A visual data with a task specification, denoted by (a, I, b, t), is employed to pretrain a
task conditional visual classifier. (c) The pretrained task conditional visual classifier is transferred to VQA and the parameters are frozen.
Attention layer and question encoder are learned from scratch with VQA dataset. The terms label and answer are used interchangeably. a ∈[0, 1]l, where the terms answer and label are used
interchangeably based on context hereafter. The classifier
formulated as a neural network with parameter θ models a
conditional distribution pθ(a|v, τ). Note that two inputs v
and τ are typically obtained by encoders vφ(·) and τη(·).
In the proposed transfer learning scenario, a task con-
ditional visual classifier is pretrained with off-the-shelf vi-
sual dataset, e.g., Visual Genome [22], and transfered to
VQA. In the pretraining stage, the parameters of the clas-
sifier and two feature encoders θ, φpre and ηpre are jointly
learned. This stage allows the task conditional visual clas-
sifier to handle diverse visual recognition tasks by learning
the task feature τ. Transfer learning to VQA is achieved by
reusing parameter θ and adapting new encoders vφvqa(·) and
τηvqa(·) to the learned task conditional visual classifier. our problem setting because they focus on predicting object
words without task specification while the task conditional
visual recognition is required for VQA.
Our problem setting is closely related to zero-shot learn-
ing [11, 17, 25, 26, 42], where out of vocabulary answers
are considered in classification. Zero-shot learning aims to
recognize objects or classes that are unseen during training.
As it aims generalization to completely unseen classes, any
exposure to zero-shot classes during training is strictly pro-
hibited [42]. On the contrary, our goal is to exploit class
labels available in external dataset. 3. Transfer Learning Framework The main goal of our work is to handle out-of-vocabulary
answers in VQA by learning visual concepts from off-the-
shelf visual dataset and transfering the concepts to VQA
for answering the questions. Inspired by the fact that VQA
can be thought of as a task conditional classification prob-
lem, where tasks are defined by questions, we introduce a
task conditional visual classifier, which generates an answer
from a visual input and a task specification, as a medium
for learning and transfering visual concepts. Figure 2 il-
lustrates the overall framework of the proposed approach.
We pretrain the task conditional visual classifier using vi-
sual dataset without questions or task specifications via un-
supervised task discovery, and adapt it to VQA models by
transferring the learned parameters. We describe the task
conditional visual classifier and how it is pretrained and
transferred to VQA in the rest of this section. 3.2. Pretraining Learning the task conditional visual classifier is naturally
formulated as the problem to maximize the following ex-
pected log likelihood: θ∗, φ∗
pre, η∗
pre = argmax
θ,φpre,ηpre
EpD

log pθ(a|vφpre(I, b), τηpre(t))

, (1)
where vφpre(I, b) is a visual feature based on an image I
and a bounding box b, and τηpre(t) is a task feature encoded
from a task specification t, a is an answer sampled from
data distribution and it satisfies a ∈A, and {θ, φpre, ηpre}
are model parameters. We obtain vφpre(I, b) using a learn-
able attention network parametrized by φpre on top of off-
the-shelf feature extractor [3], where the bounding box po-
sition b is used as a key to the attention. The optimization
in Eq. (1) requires a joint distribution, pD(a, I, b, t), which
is not accessible in the external datasets in our setting due
to missing task specifications t. Section 4 describes how
to model the joint distribution pD(a, I, b, t) with the visual 3.1. Task conditional visual classifier Task conditional visual classifier is a function taking a
visual feature v ∈Rd and a task feature τ ∈Rk and
producing a probability distribution of answers or labels",A diagram of the flow of data into an.,A detailed layout showing the process of a multi - layer task discovery system for the user and its users.,"A comprehensive technical explanation of the unsuperved task discovery model for a data driven data driven system, from the u snm / tss / tf and v / t / tg / t - t representing all details, specifications, and configurations of the system.","Figure 2. Overview of the proposed algorithm. (a) Unsupervised task discovery samples a task specification for a sampled visual data
(a, I, b), where I, b and a are an image, a bounding box and an label (answer), respectively. It leverages linguistic knowledge sources
such as visual description d and WordNet. (b) A visual data with a task specification, denoted by (a, I, b, t), is employed to pretrain a
task conditional visual classifier. (c) The pretrained task conditional visual classifier is transferred to VQA and the parameters are frozen.
Attention layer and question encoder are learned from scratch with VQA dataset. The terms label and answer are used interchangeably. a ∈[0, 1]l, where the terms answer and label are used
interchangeably based on context hereafter. The classifier
formulated as a neural network with parameter θ models a
conditional distribution pθ(a|v, τ). Note that two inputs v
and τ are typically obtained by encoders vφ(·) and τη(·).
In the proposed transfer learning scenario, a task con-
ditional visual classifier is pretrained with off-the-shelf vi-
sual dataset, e.g., Visual Genome [22], and transfered to
VQA. In the pretraining stage, the parameters of the clas-
sifier and two feature encoders θ, φpre and ηpre are jointly
learned. This stage allows the task conditional visual clas-
sifier to handle diverse visual recognition tasks by learning
the task feature τ. Transfer learning to VQA is achieved by
reusing parameter θ and adapting new encoders vφvqa(·) and
τηvqa(·) to the learned task conditional visual classifier. our problem setting because they focus on predicting object
words without task specification while the task conditional
visual recognition is required for VQA.
Our problem setting is closely related to zero-shot learn-
ing [11, 17, 25, 26, 42], where out of vocabulary answers
are considered in classification. Zero-shot learning aims to
recognize objects or classes that are unseen during training.
As it aims generalization to completely unseen classes, any
exposure to zero-shot classes during training is strictly pro-
hibited [42]. On the contrary, our goal is to exploit class
labels available in external dataset. 3. Transfer Learning Framework The main goal of our work is to handle out-of-vocabulary
answers in VQA by learning visual concepts from off-the-
shelf visual dataset and transfering the concepts to VQA
for answering the questions. Inspired by the fact that VQA
can be thought of as a task conditional classification prob-
lem, where tasks are defined by questions, we introduce a
task conditional visual classifier, which generates an answer
from a visual input and a task specification, as a medium
for learning and transfering visual concepts. Figure 2 il-
lustrates the overall framework of the proposed approach.
We pretrain the task conditional visual classifier using vi-
sual dataset without questions or task specifications via un-
supervised task discovery, and adapt it to VQA models by
transferring the learned parameters. We describe the task
conditional visual classifier and how it is pretrained and
transferred to VQA in the rest of this section. 3.2. Pretraining Learning the task conditional visual classifier is naturally
formulated as the problem to maximize the following ex-
pected log likelihood: θ∗, φ∗
pre, η∗
pre = argmax
θ,φpre,ηpre
EpD

log pθ(a|vφpre(I, b), τηpre(t))

, (1)
where vφpre(I, b) is a visual feature based on an image I
and a bounding box b, and τηpre(t) is a task feature encoded
from a task specification t, a is an answer sampled from
data distribution and it satisfies a ∈A, and {θ, φpre, ηpre}
are model parameters. We obtain vφpre(I, b) using a learn-
able attention network parametrized by φpre on top of off-
the-shelf feature extractor [3], where the bounding box po-
sition b is used as a key to the attention. The optimization
in Eq. (1) requires a joint distribution, pD(a, I, b, t), which
is not accessible in the external datasets in our setting due
to missing task specifications t. Section 4 describes how
to model the joint distribution pD(a, I, b, t) with the visual 3.1. Task conditional visual classifier Task conditional visual classifier is a function taking a
visual feature v ∈Rd and a task feature τ ∈Rk and
producing a probability distribution of answers or labels",0.7886,0.2835,0.536,2304,702,3.282,0747ffa63b7b77f2f0b758385439500a,images/2018/arxiv_0000466.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000467,Figure 467,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 2. Overview of the proposed algorithm. (a) Unsupervised task discovery samples a task specification for a sampled visual data
(a, I, b), where I, b and a are an image, a bounding box and an label (answer), respectively. It leverages linguistic knowledge sources
such as visual description d and WordNet. (b) A visual data with a task specification, denoted by (a, I, b, t), is employed to pretrain a
task conditional visual classifier. (c) The pretrained task conditional visual classifier is transferred to VQA and the parameters are frozen.
Attention layer and question encoder are learned from scratch with VQA dataset. The terms label and answer are used interchangeably.",A diagram of a computer system showing.,"A detailed layout showing the basic components of a computer architecture for the system, including the information and the description.","A comprehensive technical explanation of a typical data processing system for the computer system, including the processing process and the storage capacity for the system, as shown in this diagram from above, it is in figure 1, p4,.","Figure 2. Overview of the proposed algorithm. (a) Unsupervised task discovery samples a task specification for a sampled visual data
(a, I, b), where I, b and a are an image, a bounding box and an label (answer), respectively. It leverages linguistic knowledge sources
such as visual description d and WordNet. (b) A visual data with a task specification, denoted by (a, I, b, t), is employed to pretrain a
task conditional visual classifier. (c) The pretrained task conditional visual classifier is transferred to VQA and the parameters are frozen.
Attention layer and question encoder are learned from scratch with VQA dataset. The terms label and answer are used interchangeably.",0.81,0.2427,0.5263,1224,447,2.738,6a5102b4249b39b3a8f8f8f6b45910c5,images/2018/arxiv_0000467.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000468,Figure 468,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 3. Unsupervised task discovery with two different lin-
guistic knowledge sources. (a) For visual description, task spec-
ification is generated by replacing a visual word (label) from the
description into a blank. (b) For WordNet, task specification is a
synset from one of the labels hypernyms. We use inverted wordset
to sample a synset from an input label. See Section 4 for details. annotations and linguistic knowledge sources. 3.3. Transfer learning for VQA As illustrated in Figure 2, the proposed VQA model con-
tains a task conditional visual classifier pθ(a|v, τ). The pre-
trained visual concepts are transferred to VQA by sharing
the learned parameters θ. Then, learning a VQA model is
now formulated as learning input representations v and τ
for pθ(a|v, τ), which is given by φ∗
vqa, η∗
vqa = argmax
φvqa,ηvqa
Epvqa

log pθ(a|vφvqa(I, q), τηvqa(q))

, (2)
where vφvqa(I, q) is an encoded visual feature with an im-
age I and a question q using an attention mechanism with
parameter φvqa and a off-the-shelf feature extractor [3]. A
task feature τηvqa(q) encodes a question q using parameter
ηvqa. The joint distribution of a training dataset for VQA,
pvqa(a, I, q), is required for optimization, where answers
from the distribution satisfy a ∈A −B. We learn φvqa
and ηvqa by maximizing the likelihood of the objective in
Eq. (2) while the parameter for the pretrained task condi-
tional visual classifier θ remains fixed. Weakly supervised task regression
Utilizing a pre-
trained task conditional visual classifier for the visual recog-
nition specified by a question q requires to infer an opti-
mal task feature τ ∗
q. This requirement introduces a learn-
ing problem—task regression—that optimizes an encoder
τηvqa(q) to predict τ ∗
q correctly.
Because directly min-
imizing error E(τ ∗
q, τηvqa(q)) requires additional supervi-
sion about the tasks, we instead exploit VQA data as a",A diagram of a motorcycle with a seat.,A detailed layout showing different types of the web development process for visual discovery and visual descriptioning.,"A comprehensive technical explanation of a visual description for a motorcycle rider in a world of motorcycles and motorbikes, with the following words and descriptions below the text on the diagram, a list, a, b, a word.","Figure 3. Unsupervised task discovery with two different lin-
guistic knowledge sources. (a) For visual description, task spec-
ification is generated by replacing a visual word (label) from the
description into a blank. (b) For WordNet, task specification is a
synset from one of the labels hypernyms. We use inverted wordset
to sample a synset from an input label. See Section 4 for details. annotations and linguistic knowledge sources. 3.3. Transfer learning for VQA As illustrated in Figure 2, the proposed VQA model con-
tains a task conditional visual classifier pθ(a|v, τ). The pre-
trained visual concepts are transferred to VQA by sharing
the learned parameters θ. Then, learning a VQA model is
now formulated as learning input representations v and τ
for pθ(a|v, τ), which is given by φ∗
vqa, η∗
vqa = argmax
φvqa,ηvqa
Epvqa

log pθ(a|vφvqa(I, q), τηvqa(q))

, (2)
where vφvqa(I, q) is an encoded visual feature with an im-
age I and a question q using an attention mechanism with
parameter φvqa and a off-the-shelf feature extractor [3]. A
task feature τηvqa(q) encodes a question q using parameter
ηvqa. The joint distribution of a training dataset for VQA,
pvqa(a, I, q), is required for optimization, where answers
from the distribution satisfy a ∈A −B. We learn φvqa
and ηvqa by maximizing the likelihood of the objective in
Eq. (2) while the parameter for the pretrained task condi-
tional visual classifier θ remains fixed. Weakly supervised task regression
Utilizing a pre-
trained task conditional visual classifier for the visual recog-
nition specified by a question q requires to infer an opti-
mal task feature τ ∗
q. This requirement introduces a learn-
ing problem—task regression—that optimizes an encoder
τηvqa(q) to predict τ ∗
q correctly.
Because directly min-
imizing error E(τ ∗
q, τηvqa(q)) requires additional supervi-
sion about the tasks, we instead exploit VQA data as a",0.75,0.321,0.5355,1154,910,1.268,d895127897683531f379223c42f800f4,images/2018/arxiv_0000468.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000469,Figure 469,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 3. Unsupervised task discovery with two different lin-
guistic knowledge sources. (a) For visual description, task spec-
ification is generated by replacing a visual word (label) from the
description into a blank. (b) For WordNet, task specification is a
synset from one of the labels hypernyms. We use inverted wordset
to sample a synset from an input label. See Section 4 for details.",A diagram of a computer task with.,A detailed layout showing the visual description of a vehicle that is used for a task in a workflow.,"A comprehensive technical explanation of a visual description for a computer system, including a user and a user in a networked task environment, and a webdriverqescription, as described by the user,.","Figure 3. Unsupervised task discovery with two different lin-
guistic knowledge sources. (a) For visual description, task spec-
ification is generated by replacing a visual word (label) from the
description into a blank. (b) For WordNet, task specification is a
synset from one of the labels hypernyms. We use inverted wordset
to sample a synset from an input label. See Section 4 for details.",0.7071,0.307,0.5071,553,517,1.07,ea38b92585ab975a5314fc9c52fc2a88,images/2018/arxiv_0000469.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000470,Figure 470,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 4. Illustration of WordNet and constructed word group table. (Left) A subgraph of the WordNet [10]. Complex hierarchy of
words reveals the diverse categorization of each words. (Right) A set of words sharing common parents in the tree is grouped as a single
word group. Diverse grouping of words reveals diverse visual recognition tasks that can be defined on each word group. 4.2. Visual description hypernyms of the word in the child. In WordNet, we de-
fine a task specification tw as a synset for a node that is a
common ancestor of multiple words because a set of words
sharing a common ancestor constructs a word group and the
word group may also define a visual recognition task.
A procedure for sampling a task specification tw based
on WordNet and a visual data (a, I, b, d) is illustrated in
Figure 3b. The main idea for the procedure is to model a
task distribution conditioned on an answer p(tw|a) as a uni-
form distribution over possible word groups that the answer
belongs to, where a task specification tw is a common an-
cestor of words in an word group. Modeling the distribution
p(tw|a) requires two stages: 1) constructing a word group
table, which maps a task specification to a word group and
2) constructing an inverted word group table, which maps
an answer word to a set of task specifications. The inverted
word group table is used to retrieve a set of possible task
specifications for an answer a and the distribution p(tw|a)
is the uniform distribution over task specifications in the
set. Given the distribution p(tw|a), the joint distribution,
pD(a, I, b, tw) =
R
p(a, I, b, tw, d)dd, is given by We use Visual Genome [22] as an off-the-shelf visual
dataset, which determines a data distribution pV(a, I, b, d)
based on a set of quadruples (a, I, b, d) including visual
descriptions d. The description in this dataset is designed
to mention answer a explicitly, so that relation between the
answer and the description is clear.
To this end, we define task specification td by replacing
the answer in visual description to a special word <blank>,
which is formally denoted as td = ρ(d, a), where ρ(d, a)
is a function generating a blanked description. The sub-
script in td means that a task specification is extracted based
on a visual description. Based on this definition, joint dis-
tribution, pD(a, I, b, td) =
R
p(a, I, b, td, d)dd, with task
specification is given by p(a, I, b, td, d) = p(td|d, a)pV(a, I, b, d)
(3) where p(td|d, a) = δ

td, ρ(d, a)


is a delta function that
returns 1 if two inputs are identical and 0 otherwise. As il-
lustrated in Figure 3a, we sample data required for pretrain-
ing (a, I, b, tw) by first sampling a visual data (a, I, b, d)
and then sampling a task specification td from p(td|d, a).
This procedure results in sampling description d as well,
but we do not care about it when we pretrain the task con-
ditional visual classifier. For pretraining, we encode td into
a task feature τηpre(td) based on a gated recurrent unit [8]
because td is a sequence of words.
The main reason to use a blanked description for a task
specification is that it is effective to define a set of candidate
words. For example, a blanked description “a man is hold-
ing
” restricts candidate words for the blank to a set
of objects that can be held. Therefore, a blanked descrip-
tion can be used to determine a word group implicitly that
represents a visual recognition task. p(a, I, b, tw, d) = p(tw|a)pV(a, I, b, d).
(4) Therefore we sample a quadruple (a, I, b, d) from the vi-
sual dataset and sample a task specification tw subse-
quently. While this procedure samples a description as well,
we marginalize it out. For pretraining, we encode tw into a
task specification vector τηpre(tw) based on a word embed-
ding function that is learned from scratch.
The word group table is constructed by selecting a synset
of a node in WordNet as a task specification tw and map-
ping it to a set of words (a word group) corresponding to all
its descendants. Any word group can be defined regardless
of its level in WordNet hierarchy and the part-of-speech of
its members; the biggest word group contains all words in
WordNet and its task specification corresponds to the root
of WordNet. We illustrate the constructed word group table
in Figure 4. The inverted word group table is constructed in
a similar way to an inverted index of the word group table,
but the range of mapping is not a set of indices but a set of
task specifications. 4.3. WordNet WordNet [10] is a lexical database represented with a di-
rected acyclic graph of disambiguated word entities, called
synsets. A sample subgraph of WordNet is illustrated in
Figure 4 (left). The graph represents a hierarchical struc-
ture of words, where the parents of a node correspond to",A diagram of the structure of an.,"A detailed layout showing the field game for the baseball game, including the field games and the field.","A comprehensive technical explanation of the game in the game tree for the game, which is a set of different levels of difficulty and difficulty to play in the next level of the player ' s game ' s life ' s activity.","Figure 4. Illustration of WordNet and constructed word group table. (Left) A subgraph of the WordNet [10]. Complex hierarchy of
words reveals the diverse categorization of each words. (Right) A set of words sharing common parents in the tree is grouped as a single
word group. Diverse grouping of words reveals diverse visual recognition tasks that can be defined on each word group. 4.2. Visual description hypernyms of the word in the child. In WordNet, we de-
fine a task specification tw as a synset for a node that is a
common ancestor of multiple words because a set of words
sharing a common ancestor constructs a word group and the
word group may also define a visual recognition task.
A procedure for sampling a task specification tw based
on WordNet and a visual data (a, I, b, d) is illustrated in
Figure 3b. The main idea for the procedure is to model a
task distribution conditioned on an answer p(tw|a) as a uni-
form distribution over possible word groups that the answer
belongs to, where a task specification tw is a common an-
cestor of words in an word group. Modeling the distribution
p(tw|a) requires two stages: 1) constructing a word group
table, which maps a task specification to a word group and
2) constructing an inverted word group table, which maps
an answer word to a set of task specifications. The inverted
word group table is used to retrieve a set of possible task
specifications for an answer a and the distribution p(tw|a)
is the uniform distribution over task specifications in the
set. Given the distribution p(tw|a), the joint distribution,
pD(a, I, b, tw) =
R
p(a, I, b, tw, d)dd, is given by We use Visual Genome [22] as an off-the-shelf visual
dataset, which determines a data distribution pV(a, I, b, d)
based on a set of quadruples (a, I, b, d) including visual
descriptions d. The description in this dataset is designed
to mention answer a explicitly, so that relation between the
answer and the description is clear.
To this end, we define task specification td by replacing
the answer in visual description to a special word <blank>,
which is formally denoted as td = ρ(d, a), where ρ(d, a)
is a function generating a blanked description. The sub-
script in td means that a task specification is extracted based
on a visual description. Based on this definition, joint dis-
tribution, pD(a, I, b, td) =
R
p(a, I, b, td, d)dd, with task
specification is given by p(a, I, b, td, d) = p(td|d, a)pV(a, I, b, d)
(3) where p(td|d, a) = δ

td, ρ(d, a)


is a delta function that
returns 1 if two inputs are identical and 0 otherwise. As il-
lustrated in Figure 3a, we sample data required for pretrain-
ing (a, I, b, tw) by first sampling a visual data (a, I, b, d)
and then sampling a task specification td from p(td|d, a).
This procedure results in sampling description d as well,
but we do not care about it when we pretrain the task con-
ditional visual classifier. For pretraining, we encode td into
a task feature τηpre(td) based on a gated recurrent unit [8]
because td is a sequence of words.
The main reason to use a blanked description for a task
specification is that it is effective to define a set of candidate
words. For example, a blanked description “a man is hold-
ing
” restricts candidate words for the blank to a set
of objects that can be held. Therefore, a blanked descrip-
tion can be used to determine a word group implicitly that
represents a visual recognition task. p(a, I, b, tw, d) = p(tw|a)pV(a, I, b, d).
(4) Therefore we sample a quadruple (a, I, b, d) from the vi-
sual dataset and sample a task specification tw subse-
quently. While this procedure samples a description as well,
we marginalize it out. For pretraining, we encode tw into a
task specification vector τηpre(tw) based on a word embed-
ding function that is learned from scratch.
The word group table is constructed by selecting a synset
of a node in WordNet as a task specification tw and map-
ping it to a set of words (a word group) corresponding to all
its descendants. Any word group can be defined regardless
of its level in WordNet hierarchy and the part-of-speech of
its members; the biggest word group contains all words in
WordNet and its task specification corresponds to the root
of WordNet. We illustrate the constructed word group table
in Figure 4. The inverted word group table is constructed in
a similar way to an inverted index of the word group table,
but the range of mapping is not a set of indices but a set of
task specifications. 4.3. WordNet WordNet [10] is a lexical database represented with a di-
rected acyclic graph of disambiguated word entities, called
synsets. A sample subgraph of WordNet is illustrated in
Figure 4 (left). The graph represents a hierarchical struc-
ture of words, where the parents of a node correspond to",0.7214,0.3177,0.5195,1034,810,1.277,509a07b3a1cf1768e55bef3ff43b76d1,images/2018/arxiv_0000470.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000471,Figure 471,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 4. Illustration of WordNet and constructed word group table. (Left) A subgraph of the WordNet [10]. Complex hierarchy of
words reveals the diverse categorization of each words. (Right) A set of words sharing common parents in the tree is grouped as a single
word group. Diverse grouping of words reveals diverse visual recognition tasks that can be defined on each word group.",A diagram of the process for making a.,"A detailed layout showing the process of creating a text structure for a book, manuscript or notebook,.","A comprehensive technical explanation of a typical design for an interactive learning system, including a user ' s manual, and a user - centered design approach to the design process, as well - based approach, from the user - based.","Figure 4. Illustration of WordNet and constructed word group table. (Left) A subgraph of the WordNet [10]. Complex hierarchy of
words reveals the diverse categorization of each words. (Right) A set of words sharing common parents in the tree is grouped as a single
word group. Diverse grouping of words reveals diverse visual recognition tasks that can be defined on each word group.",0.7179,0.2512,0.4845,1224,342,3.579,eb3c8e35874a0dbd707fddb173d1855c,images/2018/arxiv_0000471.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000472,Figure 472,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 6. Data comparisons. Using visual description and WordNet shows different generalization characteristics and combining them
brings additional improvement.",A diagram of the three graphs show the.,A detailed layout showing the average and projected growth of the u s oil production from 1950 to 2011.,"A comprehensive technical explanation of the global climate and climate change in the us, from the u s, canada, and china source nasa / jps / jp / jp - 3 - jps - 1 / jpl / jp representing all details, specifications, and configurations of the.","Figure 6. Data comparisons. Using visual description and WordNet shows different generalization characteristics and combining them
brings additional improvement.",0.7179,0.1616,0.4397,1191,335,3.555,88ed4e6a9c51d67c3a0df4d12210c46d,images/2018/arxiv_0000472.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000473,Figure 473,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 5. Model comparisons. Exploiting external data with the unsupervised task discovery boosts performance of the proposed model
and separable classifier significantly while the separable classifier shows limited gains on attribute answers with large variations.",A diagram of the number of people who.,"A detailed layout showing the different types of the data in the following graphs, with the same amount of data.","A comprehensive technical explanation of the economic impact of the u s oil and gas industry in the world, from the u n d department of energy, 2008 - 2007 to 2010 - 2012 - 2013 - 13 - 08 - 11 representing all details, specifications, and.","Figure 5. Model comparisons. Exploiting external data with the unsupervised task discovery boosts performance of the proposed model
and separable classifier significantly while the separable classifier shows limited gains on attribute answers with large variations.",0.81,0.2572,0.5336,1191,335,3.555,3482e27da9e4b02beff827aae1ad7b59,images/2018/arxiv_0000473.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000474,Figure 474,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 7.
Complementary characteristics of data.
Visual
description and WordNet show complementary characteriscs in
terms of VQA score for different answers. 5. Experiments We evaluate how effectively the proposed framework
leverages the external data without questions to answer out-
of-vocabulary words in visual question answering. We com-
pare the proposed method with the baselines equipped with
idea for zero-shot image classification [11] and novel object
captioning [4, 37], which are related to the proposed prob-
lem. We also analyze the impact of the external data used
for pretraining, and visualize the mapping between ques-
tions and task specifications learned by weakly supervised",A diagram of a bar chart showing the.,A detailed layout showing the number of people who are working on their laptops in the us and canada.,"A comprehensive technical explanation of the world ' s most popular and most expensive vehicles, including the vw - vwv, vww - vr - v - v, v - vc, vx - vv, and vw vw representing all details, specifications, and configurations of the system.","Figure 7.
Complementary characteristics of data.
Visual
description and WordNet show complementary characteriscs in
terms of VQA score for different answers. 5. Experiments We evaluate how effectively the proposed framework
leverages the external data without questions to answer out-
of-vocabulary words in visual question answering. We com-
pare the proposed method with the baselines equipped with
idea for zero-shot image classification [11] and novel object
captioning [4, 37], which are related to the proposed prob-
lem. We also analyze the impact of the external data used
for pretraining, and visualize the mapping between ques-
tions and task specifications learned by weakly supervised",0.7107,0.249,0.4798,869,468,1.857,806043044cdc1a465bc50602070e8f8d,images/2018/arxiv_0000474.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000475,Figure 475,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 5. Model comparisons. Exploiting external data with the unsupervised task discovery boosts performance of the proposed model
and separable classifier significantly while the separable classifier shows limited gains on attribute answers with large variations.",A diagram of the different types of.,"A detailed layout showing the number of different types of air traffic and passenger miles per hour, from a distance.","A comprehensive technical explanation of the u s - china trade war in asia and asia, as described by the u kremalf report on the united states ' s economic impact and the world bank notes, may 2007 - 2013.","Figure 5. Model comparisons. Exploiting external data with the unsupervised task discovery boosts performance of the proposed model
and separable classifier significantly while the separable classifier shows limited gains on attribute answers with large variations.",0.75,0.1999,0.4749,1224,402,3.045,f75cff9544758af57a31d687e9d98577,images/2018/arxiv_0000475.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000476,Figure 476,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 6. Data comparisons. Using visual description and WordNet shows different generalization characteristics and combining them
brings additional improvement.",A diagram of the different types of.,"A detailed layout showing the number of active and non active users in the internet world, from 2000 to 2011.","A comprehensive technical explanation of the average and average range of data for each region, including different types of data and their corresponding datas, as well as well with examples from the following data source and data source, 2010 -.","Figure 6. Data comparisons. Using visual description and WordNet shows different generalization characteristics and combining them
brings additional improvement.",0.7393,0.2262,0.4828,1224,272,4.5,fe0df4c19aadcf2e8efe23acbd9128c0,images/2018/arxiv_0000476.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000477,Figure 477,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 7.
Complementary characteristics of data.
Visual
description and WordNet show complementary characteriscs in
terms of VQA score for different answers.",A diagram of the number of people who.,"A detailed layout showing the number of work done by the company, in the past five years and the last.","A comprehensive technical explanation of the world wide web spending by country and type of application, 2013 - 2016, based on average and selected usage of web pages of webs for each type of website content type of web contents.","Figure 7.
Complementary characteristics of data.
Visual
description and WordNet show complementary characteriscs in
terms of VQA score for different answers.",0.7143,0.2399,0.4771,553,286,1.934,9d2022c7c5749e6cec41735a83effdce,images/2018/arxiv_0000477.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000478,Figure 478,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 8. Out-of-vocabulary answers with diverse types of concepts. Green color denotes correct answers. All predicted answers are
from out-of-vocabulary answers. The proposed model successfully predicts diverse out-of-vocabulary answers depending on questions.",A diagram of the four different images.,"A detailed layout showing the different stages of a video game being played by a man and woman, including a soccer player, a tennis player, and a.","A comprehensive technical explanation of the different sports events to attend in a stadium or a stadium with pictures of people and their coach on the screen, and a person at a table with a television on the other side of the same.","Figure 8. Out-of-vocabulary answers with diverse types of concepts. Green color denotes correct answers. All predicted answers are
from out-of-vocabulary answers. The proposed model successfully predicts diverse out-of-vocabulary answers depending on questions.",0.81,0.1951,0.5026,2316,598,3.873,ebff01d3aec97f51af3d6ba32deeecc7,images/2018/arxiv_0000478.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000479,Figure 479,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 8. Out-of-vocabulary answers with diverse types of concepts. Green color denotes correct answers. All predicted answers are
from out-of-vocabulary answers. The proposed model successfully predicts diverse out-of-vocabulary answers depending on questions.",A diagram of the different types of.,"A detailed layout showing four different photos of a person in a room with a tv and other people that shows key features, attributes, and data.","A comprehensive technical explanation of the four different sports videos in this picture, including a tv and a football game on the screen, which is also an image of a person with a man standing in front of a woman in a soccer player.","Figure 8. Out-of-vocabulary answers with diverse types of concepts. Green color denotes correct answers. All predicted answers are
from out-of-vocabulary answers. The proposed model successfully predicts diverse out-of-vocabulary answers depending on questions.",0.81,0.299,0.5545,1224,400,3.06,00d36bfbd90eef93aadd1579c138d050,images/2018/arxiv_0000479.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000480,Figure 480,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 9. Combining knowledge from VQA and external visual data. Evaluation results on a test set containing both out-of-vocabulary
answers and trained answers. The proposed model shows relatively lower performance on trained answers but significantly better perfor-
mance on out-of-vocabulary answers. In total, the proposed model shows the best performance. periment, the test set also contains training answers which
include logical answers, numbers and visual words. The
list of out-of-vocabulary answers are identical to that of the
main experiment. Among 172,681 test questions, 103,013
questions can be answered only with the training answers.
To combine knowledges from VQA dataset and exter-
nal visual data, we learn a VQA model with two task con-
ditional visual classifiers; we fine-tune one classifier for
adapting answers requiring visual reasoning (i.e., numbers
and yes/no) and fix the other classifier for visual answers in-
cluding out-of-vocabulary answers. After training the VQA
model, we combine two logits by element-wise sum and
pick the answer with the highest score in the inference.
The results are presented in Figure 9. Models in each
method are trained with 6 different random seeds and their
mean and standard deviation are plotted. Overall, the pro-
posed model performs the best. While the standard VQA
model achieves the best performance for training answers,
it fails to predict any out-of-vocabulary answers. The an-
swer embedding baseline somewhat generalizes to out-of-
vocabulary answers, but constraints in the answer embed-
ding degrade its performance on answers in training set. mentary characteristics of models are illustrated in Figure 7,
where we visualize average VQA scores for 20 answers. Qualitative results
Figure 8 shows examples of predicted
answers from the proposed model. The proposed model
correctly predicts out-of-vocabulary answers for questions
asking diverse visual concepts such as type of flooring, ma-
terial, type of sport and brand. Weakly supervised task regression
Given that task spec-
ifications extracted from WordNet models diverse visual
recognition tasks, matching them to relevant questions is
useful for categorization of VQA data and model interpre-
tation. As we learn VQA models by task regression, this
matching can be performed by comparing the encoded task
feature from a question τηvqa(q) and the vector from a task
specification τηpre(tw). For each τηpre(tw), we sorted ques-
tions in a descending order of dot product similarity be-
tween τηpre(tw) and τηvqa(q). In the sorted question list, the
most similar questions are visualized in Table 1. The visu-
alization shows that the weakly supervised task regression
successfully trains a question encoder that match a question
to a relevant task feature. 6. Conclusion 5.4. Combining knowledge learned by VQA While we focus on learning visual concepts from exter-
nal visual data, VQA dataset is still a valuable source of
learning diverse knowledges. Especially, some answers are
not visual words and require visual reasoning. For exam-
ple, yes and no are one of the most frequent answers in the
VQA dataset [5] but it is not straightforward to learn these
answers only with the external visual data. Therefore, we
consider combining knowledge learned from VQA dataset
and from external visual data.
We construct a split of the VQA dataset consisting of
405,228 training, 37,031 validation, and 172,681 test ques-
tions. The training and validation set do not contain any
out-of-vocabulary answers and test set contains out-of-
vocabulary answers.
However, contrary to the main ex- We present a transfer learning approach for visual ques-
tion answering with out-of-vocabulary answers. We pre-
train a task conditional visual classifier with off-the-shelf
visual and linguistic data based on unsupervised task dis-
covery.
The pretrained task conditional visual classifier
is transferred to VQA adaptively.
The experimental re-
sults show that exploiting external visual and linguistic data
boosts performance in the proposed setting and training
with unsupervised task discovery is important to model in-
teraction between visual features and task specifications. Acknowledgments
This research was partly supported by
Kakao and Kakao Brain and Korean ICT R&D program of
the MSIP/IITP grant [2016-0-00563, 2017-0-01778].",A diagram of the different types of.,A detailed layout showing the differences in the number of different types of internet traffic and the number one that.,"A comprehensive technical explanation of the average and expected exchange rate for the fed and us stocks, as described by the fed stock market index and fedexor indexr on the feds, and fed and feds and fed.","Figure 9. Combining knowledge from VQA and external visual data. Evaluation results on a test set containing both out-of-vocabulary
answers and trained answers. The proposed model shows relatively lower performance on trained answers but significantly better perfor-
mance on out-of-vocabulary answers. In total, the proposed model shows the best performance. periment, the test set also contains training answers which
include logical answers, numbers and visual words. The
list of out-of-vocabulary answers are identical to that of the
main experiment. Among 172,681 test questions, 103,013
questions can be answered only with the training answers.
To combine knowledges from VQA dataset and exter-
nal visual data, we learn a VQA model with two task con-
ditional visual classifiers; we fine-tune one classifier for
adapting answers requiring visual reasoning (i.e., numbers
and yes/no) and fix the other classifier for visual answers in-
cluding out-of-vocabulary answers. After training the VQA
model, we combine two logits by element-wise sum and
pick the answer with the highest score in the inference.
The results are presented in Figure 9. Models in each
method are trained with 6 different random seeds and their
mean and standard deviation are plotted. Overall, the pro-
posed model performs the best. While the standard VQA
model achieves the best performance for training answers,
it fails to predict any out-of-vocabulary answers. The an-
swer embedding baseline somewhat generalizes to out-of-
vocabulary answers, but constraints in the answer embed-
ding degrade its performance on answers in training set. mentary characteristics of models are illustrated in Figure 7,
where we visualize average VQA scores for 20 answers. Qualitative results
Figure 8 shows examples of predicted
answers from the proposed model. The proposed model
correctly predicts out-of-vocabulary answers for questions
asking diverse visual concepts such as type of flooring, ma-
terial, type of sport and brand. Weakly supervised task regression
Given that task spec-
ifications extracted from WordNet models diverse visual
recognition tasks, matching them to relevant questions is
useful for categorization of VQA data and model interpre-
tation. As we learn VQA models by task regression, this
matching can be performed by comparing the encoded task
feature from a question τηvqa(q) and the vector from a task
specification τηpre(tw). For each τηpre(tw), we sorted ques-
tions in a descending order of dot product similarity be-
tween τηpre(tw) and τηvqa(q). In the sorted question list, the
most similar questions are visualized in Table 1. The visu-
alization shows that the weakly supervised task regression
successfully trains a question encoder that match a question
to a relevant task feature. 6. Conclusion 5.4. Combining knowledge learned by VQA While we focus on learning visual concepts from exter-
nal visual data, VQA dataset is still a valuable source of
learning diverse knowledges. Especially, some answers are
not visual words and require visual reasoning. For exam-
ple, yes and no are one of the most frequent answers in the
VQA dataset [5] but it is not straightforward to learn these
answers only with the external visual data. Therefore, we
consider combining knowledge learned from VQA dataset
and from external visual data.
We construct a split of the VQA dataset consisting of
405,228 training, 37,031 validation, and 172,681 test ques-
tions. The training and validation set do not contain any
out-of-vocabulary answers and test set contains out-of-
vocabulary answers.
However, contrary to the main ex- We present a transfer learning approach for visual ques-
tion answering with out-of-vocabulary answers. We pre-
train a task conditional visual classifier with off-the-shelf
visual and linguistic data based on unsupervised task dis-
covery.
The pretrained task conditional visual classifier
is transferred to VQA adaptively.
The experimental re-
sults show that exploiting external visual and linguistic data
boosts performance in the proposed setting and training
with unsupervised task discovery is important to model in-
teraction between visual features and task specifications. Acknowledgments
This research was partly supported by
Kakao and Kakao Brain and Korean ICT R&D program of
the MSIP/IITP grant [2016-0-00563, 2017-0-01778].",0.75,0.2065,0.4783,1191,335,3.555,06e5a0826acd82957485f8c1361fb1fc,images/2018/arxiv_0000480.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000481,Figure 481,scientific_figure,Transfer Learning via Unsupervised Task Discovery for Visual Question Answering,arXiv Research Authors,1810.02358v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Transfer Learning via Unsupervised Task Discovery for Visual Question Answering. arXiv:1810.02358v2,"Figure 9. Combining knowledge from VQA and external visual data. Evaluation results on a test set containing both out-of-vocabulary
answers and trained answers. The proposed model shows relatively lower performance on trained answers but significantly better perfor-
mance on out-of-vocabulary answers. In total, the proposed model shows the best performance.",A diagram of a line graph shows the.,A detailed layout showing the number of companies in each country and the number on each side of them.,"A comprehensive technical explanation of the cost of the internet usage in the us, and how much is it worth to use? infographics com / @ tm / tk / tmex / twc / t representing all details, specifications, and configurations of the system components.","Figure 9. Combining knowledge from VQA and external visual data. Evaluation results on a test set containing both out-of-vocabulary
answers and trained answers. The proposed model shows relatively lower performance on trained answers but significantly better perfor-
mance on out-of-vocabulary answers. In total, the proposed model shows the best performance.",0.7143,0.1545,0.4344,1224,408,3.0,f964ba028b3304e0117f74c68ce6114e,images/2018/arxiv_0000481.png,https://arxiv.org/pdf/1810.02358v2.pdf
arxiv_0000482,Figure 482,scientific_figure,Embedding Geographic Locations for Modelling the Natural Environment using Flickr Tags and Structured Data,arXiv Research Authors,1810.12091v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Embedding Geographic Locations for Modelling the Natural Environment using Flickr Tags and Structured Data. arXiv:1810.12091v1,"Fig. 1: Comparison between the perfor-
mance of the GloVe and bag-of-words
models for predicting scenicness, as a func-
tion of the number of tag occurrences at
the considered locations.",A diagram of the number of cases of.,A detailed layout showing the distribution of the number of tags in the above row to show that each tag is the same size as the next row.,"A comprehensive technical explanation of the growth of bow and bow prostensors in the united states, by age and gender, from 1900 to 2000 to 2010, from the present at different ages, as well as per 1,.","Fig. 1: Comparison between the perfor-
mance of the GloVe and bag-of-words
models for predicting scenicness, as a func-
tion of the number of tag occurrences at
the considered locations.",0.81,0.2913,0.5507,482,375,1.285,40c27e2f1bc9bd898f90dc6c59a9f5cf,images/2018/arxiv_0000482.png,https://arxiv.org/pdf/1810.12091v1.pdf
arxiv_0000483,Figure 483,scientific_figure,Embedding Geographic Locations for Modelling the Natural Environment using Flickr Tags and Structured Data,arXiv Research Authors,1810.12091v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Embedding Geographic Locations for Modelling the Natural Environment using Flickr Tags and Structured Data. arXiv:1810.12091v1,"Fig. 1: Comparison between the perfor-
mance of the GloVe and bag-of-words
models for predicting scenicness, as a func-
tion of the number of tag occurrences at
the considered locations.",A diagram of the number of people who.,"A detailed layout showing the average and range of t - shirts sold by the club in the past that shows key features, attributes, and data points in.","A comprehensive technical explanation of the average time of a football game in the united states and mexico, by league and region, from 1900 to 2000, 2000 - 2000 to 2012, based on a bar chart of age and below.","Fig. 1: Comparison between the perfor-
mance of the GloVe and bag-of-words
models for predicting scenicness, as a func-
tion of the number of tag occurrences at
the considered locations.",0.81,0.2371,0.5236,448,493,0.909,e521de0cd28554dcf0f033565ea65ba8,images/2018/arxiv_0000483.png,https://arxiv.org/pdf/1810.12091v1.pdf
arxiv_0000484,Figure 484,scientific_figure,Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning,arXiv Research Authors,1812.00898v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning. arXiv:1812.00898v1,"Figure 1: We build agents that can generate programs for diverse scenes conditioned on a given
symbolic instruction. (a) Conditioned on a given instruction (e.g., “There is a small sphere”.), the
agent learns to generate programs for 3D scenes such that all scenes have a small sphere. We
can see that the generated scenes show diversity in unspecified attributes (color and location of the
sphere). (b) Conditioned on a simple instruction (e.g., “Draw zero”, “Paint 1”, etc.), the agent learns
to generate programs that draw the corresponding MNIST digits. We can see that the generated
drawings of a given label show diversity in style. (c) and (d): Results corresponding to (a) and (b)
respectively for the fixed pixel-based reward function baseline. We can see that the scenes / drawings
rendered from the generated programs are either not consistent with the input instruction (c) or do
not show sufficient diversity (d).",A diagram of a number of different.,"A detailed layout showing the different types of objects in a game, including the letters and numbers on each side.","A comprehensive technical explanation of the different types of blocks in the game, including the different shapes and sizes that are used in each type of block set of legos and pieces of blocks, as well as shown in this set.","Figure 1: We build agents that can generate programs for diverse scenes conditioned on a given
symbolic instruction. (a) Conditioned on a given instruction (e.g., “There is a small sphere”.), the
agent learns to generate programs for 3D scenes such that all scenes have a small sphere. We
can see that the generated scenes show diversity in unspecified attributes (color and location of the
sphere). (b) Conditioned on a simple instruction (e.g., “Draw zero”, “Paint 1”, etc.), the agent learns
to generate programs that draw the corresponding MNIST digits. We can see that the generated
drawings of a given label show diversity in style. (c) and (d): Results corresponding to (a) and (b)
respectively for the fixed pixel-based reward function baseline. We can see that the scenes / drawings
rendered from the generated programs are either not consistent with the input instruction (c) or do
not show sufficient diversity (d).",0.75,0.285,0.5175,1456,695,2.095,15682b6bfb1ca092d900b31d9130c8ee,images/2018/arxiv_0000484.png,https://arxiv.org/pdf/1812.00898v1.pdf
arxiv_0000485,Figure 485,scientific_figure,Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning,arXiv Research Authors,1812.00898v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning. arXiv:1812.00898v1,respectively for the fixed pixel-based reward function baseline. We can see that the scenes / drawings rendered from the generated programs are either not consistent with the input instruction,A diagram of a small sphere and a.,"A detailed layout showing several different shapes of objects in a series of three rows, each with different color.","A comprehensive technical explanation of there is a small sphere, there is large cylinder, there are a yellow cube, there ' s a small square and a smaller cube, and a small cube, with a smaller one in the same color.",respectively for the fixed pixel-based reward function baseline. We can see that the scenes / drawings rendered from the generated programs are either not consistent with the input instruction,0.75,0.2627,0.5063,2342,1104,2.121,fa9035fd69298ebdb67c68bf49863423,images/2018/arxiv_0000485.png,https://arxiv.org/pdf/1812.00898v1.pdf
arxiv_0000486,Figure 486,scientific_figure,Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning,arXiv Research Authors,1812.00898v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning. arXiv:1812.00898v1,and,A diagram of the number of letters and.,"A detailed layout showing the numbers and symbols of a handwritten number line with a pen, ink, and pencil.","A comprehensive technical explanation of the handwriting of the first person to write letters in cursive writing, including the number and capital letters of the alphabets of the letter fo the curs, and the first handwritten curs.",and,0.7321,0.2942,0.5131,571,768,0.743,1de4a722c922f20c30a78ea553319e24,images/2018/arxiv_0000486.png,https://arxiv.org/pdf/1812.00898v1.pdf
arxiv_0000487,Figure 487,scientific_figure,Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning,arXiv Research Authors,1812.00898v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning. arXiv:1812.00898v1,or do not show sufficient diversity,A diagram of the numbers for each.,"A detailed layout showing the numbers of the numbers in each row, including the number and the time.","A comprehensive technical explanation of the numbers in the game, including one number, and one number and two numbers on each side of the screen that shows the scoreboard, on the screen, with a set of the number and the score.",or do not show sufficient diversity,0.7071,0.3222,0.5146,563,767,0.734,4a62008bd77165b024ad87cfacba2c26,images/2018/arxiv_0000487.png,https://arxiv.org/pdf/1812.00898v1.pdf
arxiv_0000488,Figure 488,scientific_figure,Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning,arXiv Research Authors,1812.00898v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning. arXiv:1812.00898v1,"Figure 1: We build agents that can generate programs for diverse scenes conditioned on a given
symbolic instruction. (a) Conditioned on a given instruction (e.g., “There is a small sphere”.), the
agent learns to generate programs for 3D scenes such that all scenes have a small sphere. We
can see that the generated scenes show diversity in unspecified attributes (color and location of the
sphere). (b) Conditioned on a simple instruction (e.g., “Draw zero”, “Paint 1”, etc.), the agent learns
to generate programs that draw the corresponding MNIST digits. We can see that the generated
drawings of a given label show diversity in style. (c) and (d): Results corresponding to (a) and (b)
respectively for the fixed pixel-based reward function baseline. We can see that the scenes / drawings
rendered from the generated programs are either not consistent with the input instruction (c) or do
not show sufficient diversity (d).",A diagram of the different types of.,"A detailed layout showing the different types of the letters in each letter, and their corresponding numbers and symbols.","A comprehensive technical explanation of the structure of a cell phone number chart, including the number of each cell phone, and the number and location of the number, as well as it is the number that is, there.","Figure 1: We build agents that can generate programs for diverse scenes conditioned on a given
symbolic instruction. (a) Conditioned on a given instruction (e.g., “There is a small sphere”.), the
agent learns to generate programs for 3D scenes such that all scenes have a small sphere. We
can see that the generated scenes show diversity in unspecified attributes (color and location of the
sphere). (b) Conditioned on a simple instruction (e.g., “Draw zero”, “Paint 1”, etc.), the agent learns
to generate programs that draw the corresponding MNIST digits. We can see that the generated
drawings of a given label show diversity in style. (c) and (d): Results corresponding to (a) and (b)
respectively for the fixed pixel-based reward function baseline. We can see that the scenes / drawings
rendered from the generated programs are either not consistent with the input instruction (c) or do
not show sufficient diversity (d).",0.75,0.2758,0.5129,674,540,1.248,84dcc95a79f0d2d982f3b6bf506b3a1a,images/2018/arxiv_0000488.png,https://arxiv.org/pdf/1812.00898v1.pdf
arxiv_0000489,Figure 489,scientific_figure,Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning,arXiv Research Authors,1812.00898v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning. arXiv:1812.00898v1,"Figure 2: An overview of our approach. Given an instruction the generator (policy network) outputs
a program which is rendered by a non-differentiable (denoted by dashed arrows) renderer into an im-
age. This process repeats for a fixed number of time steps. The final image is fed to a discriminator,
along with the instruction (the same that is given to the generator), that produces a reward.",A diagram of the process for a new.,"A detailed layout showing the different types of the process and process of a project, including the proposed development.","A comprehensive technical explanation of the process for the development of a multi - layer, modular system that includes multiple physical elements and a multilinguise model of the system for each other parts of the workflowermr.","Figure 2: An overview of our approach. Given an instruction the generator (policy network) outputs
a program which is rendered by a non-differentiable (denoted by dashed arrows) renderer into an im-
age. This process repeats for a fixed number of time steps. The final image is fed to a discriminator,
along with the instruction (the same that is given to the generator), that produces a reward.",0.75,0.2812,0.5156,714,222,3.216,650e5f3671abc5be0c743e0a82e88b94,images/2018/arxiv_0000489.png,https://arxiv.org/pdf/1812.00898v1.pdf
arxiv_0000490,Figure 490,scientific_figure,Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning,arXiv Research Authors,1812.00898v1,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning. arXiv:1812.00898v1,"Figure 3: (a) Discriminator’s architecture. It takes as inputs the instruction and an image (either
the image generated by the agent or a goal image from the dataset), learns a joint embedding, and
outputs a scalar score evaluating the ‘realness’ of the input pair. (b) Policy Network’s architecture,
unrolled over time. At each time step, it takes – an instruction, the rendered image from the previous
time step (blank canvas initially) and the previous action, learns a joint embedding and passes it to
an LSTM whose output is fed to a decoder which samples the actions.",A diagram of the process for a.,A detailed layout showing the process of the proposed 3d printing system for the 3d printer and the three components that could be used to produce.,"A comprehensive technical explanation of the hardware for the first generation of the qms system, including the same components as the two other systems in the diagram below, are shown below, and above the following diagrams, the picture below.","Figure 3: (a) Discriminator’s architecture. It takes as inputs the instruction and an image (either
the image generated by the agent or a goal image from the dataset), learns a joint embedding, and
outputs a scalar score evaluating the ‘realness’ of the input pair. (b) Policy Network’s architecture,
unrolled over time. At each time step, it takes – an instruction, the rendered image from the previous
time step (blank canvas initially) and the previous action, learns a joint embedding and passes it to
an LSTM whose output is fed to a decoder which samples the actions.",0.75,0.2652,0.5076,754,382,1.974,7182d0b22554527e9d53e490b26f28c4,images/2018/arxiv_0000490.png,https://arxiv.org/pdf/1812.00898v1.pdf
arxiv_0000491,Figure 491,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Agent path
(a)
(b)
(c)
(d)
(e)
Figure 1: An example run in an unseen environment. (a) A bird-eye view of the environment annotated with the agent’s path.
The agent observes the environment only through a first-person view. (b) A requester (wearing a hat) asks the agent to “find
a towel in the kitchen”. Two towels (pink circle) are in front of the agent but the room is labeled as a “bathroom”. The agent
ignores them without being given the room label. (c) The agent escapes the bathroom but runs into an unfamiliar region.
Sensing that it is lost, the agent signals the advisor (with mustache) for help. The advisor responds with an “easier” low-level
subgoal “turn 60 degrees right, go forward, turn left”. (d) After executing the subgoal, the agent is closer to the kitchen but is
still confused. It thus requests help one more time. After making this request, the agent has exhausted its request budget and
can only rely on its own. (e) Executing the second subgoal helps the agent see the target towel (cyan circle). It successfully
walks to the goal without further assistance. A video demo is at https://youtu.be/Vp6C29qTKQ0.",A diagram of a house with multiple.,"A detailed layout showing the living and dining areas of an apartment building in the city of london, england.","A comprehensive technical explanation of a typical apartment layout in 3d, including all the bedrooms and bathrooms, and the kitchen, living room, and dining area, and bedroom areas, and bathroom and bedroom, all in one of which are separated.","Agent path
(a)
(b)
(c)
(d)
(e)
Figure 1: An example run in an unseen environment. (a) A bird-eye view of the environment annotated with the agent’s path.
The agent observes the environment only through a first-person view. (b) A requester (wearing a hat) asks the agent to “find
a towel in the kitchen”. Two towels (pink circle) are in front of the agent but the room is labeled as a “bathroom”. The agent
ignores them without being given the room label. (c) The agent escapes the bathroom but runs into an unfamiliar region.
Sensing that it is lost, the agent signals the advisor (with mustache) for help. The advisor responds with an “easier” low-level
subgoal “turn 60 degrees right, go forward, turn left”. (d) After executing the subgoal, the agent is closer to the kitchen but is
still confused. It thus requests help one more time. After making this request, the agent has exhausted its request budget and
can only rely on its own. (e) Executing the second subgoal helps the agent see the target towel (cyan circle). It successfully
walks to the goal without further assistance. A video demo is at https://youtu.be/Vp6C29qTKQ0.",0.7429,0.2955,0.5192,533,441,1.209,154bacedb597f7a62020ad06c299058d,images/2018/arxiv_0000491.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000492,Figure 492,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Figure 2: Comparison between I3L trained with behavior
cloning under interventions (I3L-BCUI), imitation learning
(IL), and behavior cloning (BC) at training time (left) and
test time (right). Gray dots represent states and arrows rep-
resent actions. Bounding boxes of different colors represent
different environments.",A diagram of the three steps in the.,"A detailed layout showing how the behavior of an object works and how it can be used to teach that shows key features, attributes, and data points.","A comprehensive technical explanation of the training process for a team of engineers and technicians, including the initial steps in the training, and the initial stages of the work taking place of the teamwork on the next step,.","Figure 2: Comparison between I3L trained with behavior
cloning under interventions (I3L-BCUI), imitation learning
(IL), and behavior cloning (BC) at training time (left) and
test time (right). Gray dots represent states and arrows rep-
resent actions. Bounding boxes of different colors represent
different environments.",0.81,0.2674,0.5387,516,380,1.358,3e99bdeefdee9bd7a16eb3b858933f0e,images/2018/arxiv_0000492.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000493,Figure 493,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Figure 3: Two decoding passes of the navigation module.
(a) The first decoding pass computes the tentative naviga-
tion distribution, which is used as a feature for computing
the help-requesting distribution. (b) The second pass com-
putes the final navigation distribution.",A diagram of the sequence of a memory.,"A detailed layout showing the different types of the organization in the system, including an active memory,.","A comprehensive technical explanation of the algorithm for a fixed memory and a dedicated memory in the same system, each with an array of memory paths to it ' s own sequences, a, b, c, d, e representing all details, specifications, and.","Figure 3: Two decoding passes of the navigation module.
(a) The first decoding pass computes the tentative naviga-
tion distribution, which is used as a feature for computing
the help-requesting distribution. (b) The second pass com-
putes the final navigation distribution.",0.7393,0.2523,0.4958,490,242,2.025,970620f7115af6b0bb99bdaad83398cc,images/2018/arxiv_0000493.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000494,Figure 494,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,Figure 4: Top 20 most common objects in the ASKNAV dataset.,A diagram of the number of people in.,A detailed layout showing the number of people who are visiting the uk in each country of the world.,"A comprehensive technical explanation of the number of covidtors in the uk, by age and gender, by type of population, 2010 - 2016, by category, and region, as explained with data, 2013 - q1 representing all details, specifications, and.",Figure 4: Top 20 most common objects in the ASKNAV dataset.,0.7071,0.2461,0.4766,590,358,1.648,671a44f7bc9224b23b969ca6f378fb3a,images/2018/arxiv_0000494.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000495,Figure 495,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,Figure 5: Top 20 most common start rooms in the ASKNAV dataset.,A diagram of the number of people who.,"A detailed layout showing the average and most time of day to day traffic in the city of london that shows key features, attributes, and data points.","A comprehensive technical explanation of the average and medianised time for a t2n - t4n - v1n - m9n - 1n - s1 - t2 - c3s - 3 - h representing all details, specifications, and configurations of the system components in full context with annotated.",Figure 5: Top 20 most common start rooms in the ASKNAV dataset.,0.81,0.237,0.5235,623,354,1.76,2bc3f0f49045fd871c0ad2fa854df64c,images/2018/arxiv_0000495.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000496,Figure 496,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,Figure 6: Top 20 most common goal rooms in the ASKNAV dataset.,A diagram of the number of people who.,A detailed layout showing the number of people who have been given a different type of data in the past.,"A comprehensive technical explanation of the number of people in the uk who are using twitter to get more followers and get more traffic from their website posts and videos on twitter commund on twitter, but not just one is still.",Figure 6: Top 20 most common goal rooms in the ASKNAV dataset.,0.7214,0.2514,0.4864,623,354,1.76,fcbda6a2997f6fef35c45ede30bb86d3,images/2018/arxiv_0000496.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000497,Figure 497,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Figure 7:
Distribution of path lengths in the ASKNAV
dataset. Paths are computed by the shortest-path navigation
teacher.",A diagram of the different types of.,A detailed layout showing the number of different types of t - shirts and how they are used to make them.,"A comprehensive technical explanation of the number of tests performed by the team in the study, including the results of each test, and their results on the result, on the results, in the results and on the test,.","Figure 7:
Distribution of path lengths in the ASKNAV
dataset. Paths are computed by the shortest-path navigation
teacher.",0.725,0.1899,0.4575,513,188,2.729,995ae77091de877646d0e354660841b5,images/2018/arxiv_0000497.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000498,Figure 498,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Figure 8: Fraction of help requests made over (normalized)
time.",A diagram of a graph showing the.,"A detailed layout showing the normalized time of test results for the two test - set tests that shows key features, attributes, and data points in.","A comprehensive technical explanation of test - driven tests for the normalized time and the normalised time in the test, from u s t c l e e v, p m t, a t e t e, s representing all details, specifications, and configurations of the system.","Figure 8: Fraction of help requests made over (normalized)
time.",0.87,0.292,0.581,466,464,1.004,c804b23b67bbf324f10855b7051804d4,images/2018/arxiv_0000498.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000499,Figure 499,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Figure 9: Success rate versus number of actions taken by
(a) the navigation teacher and (b) the agent. Error bars are
95% confidence intervals.",A diagram of the number of actions.,A detailed layout showing the various levels of activity in the experiment and how they can be achieved by each individual.,"A comprehensive technical explanation of the study of the number of actions in each of the following groups, by the number and number of action in each group of actions, by team members, and the team members of the team,.","Figure 9: Success rate versus number of actions taken by
(a) the navigation teacher and (b) the agent. Error bars are
95% confidence intervals.",0.75,0.212,0.481,510,268,1.903,34b7e56165a7a553ce56d38665f47368,images/2018/arxiv_0000499.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000500,Figure 500,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Figure 10: Top five objects with highest and lowest aver-
age success rates in (a) TEST SEEN and (b) TEST UNSEEN.
Numbers in parentheses are object frequencies. Only ob-
jects appearing more than 50 times are included. Error bars
are 95% confidence intervals.",A diagram of the average and depicting.,A detailed layout showing the number of people who are using different languages in the language of the language.,"A comprehensive technical explanation of the data visual for the graph of a set of data visual data visual to a bar chart with different data visual bars and numbers in red and blue and orange bars on white bars, from left to right.","Figure 10: Top five objects with highest and lowest aver-
age success rates in (a) TEST SEEN and (b) TEST UNSEEN.
Numbers in parentheses are object frequencies. Only ob-
jects appearing more than 50 times are included. Error bars
are 95% confidence intervals.",0.75,0.2324,0.4912,509,267,1.906,d88b8b6a33970f1ba68950c2b8af3009,images/2018/arxiv_0000500.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000501,Figure 501,scientific_figure,Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention,arXiv Research Authors,1812.04155v4,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Vision-based Navigation with Language-based Assistance via Imitation Learning with Indirect Intervention. arXiv:1812.04155v4,"Figure 11: Top five goal rooms with highest and lowest av-
erage success rates in (a) TEST SEEN and (b) TEST UN-
SEEN. Numbers in parentheses are room frequencies. Only
rooms appearing more than 50 times are included. Error
bars are 95% confidence intervals.",A diagram of the number of people with.,A detailed layout showing the differences in the number of different types of people who have visited each country.,"A comprehensive technical explanation of the effect of a number of different types of data on an image, including data visual and data visual data visual, from a to a chart and data driven to a bar graph, and tableau.","Figure 11: Top five goal rooms with highest and lowest av-
erage success rates in (a) TEST SEEN and (b) TEST UN-
SEEN. Numbers in parentheses are room frequencies. Only
rooms appearing more than 50 times are included. Error
bars are 95% confidence intervals.",0.75,0.2151,0.4826,509,268,1.899,e91084845e88dc0d85fa308307a55f43,images/2018/arxiv_0000501.png,https://arxiv.org/pdf/1812.04155v4.pdf
arxiv_0000502,Figure 502,scientific_figure,Deep Anomaly Detection with Outlier Exposure,arXiv Research Authors,1812.04606v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Anomaly Detection with Outlier Exposure. arXiv:1812.04606v3,"Figure 1: ROC curve with Tiny Im-
ageNet (Din) and Textures (Dtest
out ).",A diagram of the curves in the curve.,A detailed layout showing the curves of the text and the slope of the image that indicates the size of the texture.,"A comprehensive technical explanation of rcc curve of textures versus tiny imageset png, rgb and msp for random random random data, rgs, rg png and rgs and rgp, rgp representing all details, specifications, and configurations of the system.","Figure 1: ROC curve with Tiny Im-
ageNet (Din) and Textures (Dtest
out ).",0.75,0.2894,0.5197,319,333,0.958,11909fb4acfe091554fa0403070aaa40,images/2018/arxiv_0000502.png,https://arxiv.org/pdf/1812.04606v3.pdf
arxiv_0000503,Figure 503,scientific_figure,Deep Anomaly Detection with Outlier Exposure,arXiv Research Authors,1812.04606v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Anomaly Detection with Outlier Exposure. arXiv:1812.04606v3,"Figure 2: OOD scores from
PixelCNN++ on images from
CIFAR-10 and SVHN.",A diagram of the results of the horse.,"A detailed layout showing the results of the three horses ' results from a horse ' s data sheet that shows key features, attributes, and data points.","A comprehensive technical explanation of the impact of the horse ' s performance in the field of competition, from the national horse magazine, july 2012, page 2, p2, page 6, p3, 2014, png representing all details, specifications, and.","Figure 2: OOD scores from
PixelCNN++ on images from
CIFAR-10 and SVHN.",0.87,0.2368,0.5534,298,510,0.584,ee9676872d74acb1942d52ac31bf4dcf,images/2018/arxiv_0000503.png,https://arxiv.org/pdf/1812.04606v3.pdf
arxiv_0000504,Figure 504,scientific_figure,Deep Anomaly Detection with Outlier Exposure,arXiv Research Authors,1812.04606v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Anomaly Detection with Outlier Exposure. arXiv:1812.04606v3,"Figure 3: Root Mean Square Calibration Error
values with temperature tuning and temperature
tuning + OE across various datasets.",A diagram of a bar graph showing the.,"A detailed layout showing the number of different types of temperatures for each type of experiment, with different levels.","A comprehensive technical explanation of the calibration error with outer exposure and temperature tuning of various instruments in two different types of filters, including the following model and the experiment results from the above the datas.","Figure 3: Root Mean Square Calibration Error
values with temperature tuning and temperature
tuning + OE across various datasets.",0.75,0.2116,0.4808,421,349,1.206,76c5dcb97e52dbedf6daff4cc57c97e0,images/2018/arxiv_0000504.png,https://arxiv.org/pdf/1812.04606v3.pdf
arxiv_0000505,Figure 505,scientific_figure,Deep Anomaly Detection with Outlier Exposure,arXiv Research Authors,1812.04606v3,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Deep Anomaly Detection with Outlier Exposure. arXiv:1812.04606v3,"Figure 4: ROC curves with Tiny ImageNet as Din and Textures, Places365, LSUN, and ImageNet as
Dtest
out . Figures show the curves corresponding to the maximum softmax probability (MSP) baseline
detector and the MSP detector with Outlier Exposure (OE).",A diagram of a series of four graphs.,"A detailed layout showing the different types of images for each image, and their corresponding properties, as well as.","A comprehensive technical explanation of the results of the multiple real - time images using the rasl and the digital imager software, part 1 of the same image set is shown below in figure 5, and the image size.","Figure 4: ROC curves with Tiny ImageNet as Din and Textures, Places365, LSUN, and ImageNet as
Dtest
out . Figures show the curves corresponding to the maximum softmax probability (MSP) baseline
detector and the MSP detector with Outlier Exposure (OE).",0.75,0.2605,0.5052,833,466,1.788,bc9b796b4b20838790d07110fe929a2a,images/2018/arxiv_0000505.png,https://arxiv.org/pdf/1812.04606v3.pdf
arxiv_0000506,Figure 506,scientific_figure,Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities,arXiv Research Authors,1812.07809v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities. arXiv:1812.07809v2,"Figure 1: Learning robust joint representations via multimodal
cyclic translations. Top: cyclic translations from a source modality
(language) to a target modality (visual). Bottom: the representation
learned between language and vision are further translated into the
acoustic modality, forming the final joint representation. In both
cases, the joint representation is then used for sentiment prediction.",A diagram of three different types of.,A detailed layout showing the effects of a virtual translation and a visual reconstruction for a patient ' s speech.,"A comprehensive technical explanation of the brain and spinal cycle of the human body, presented in this graphic from the department of health and medicine, fall 2006 - page 4 - page 3 - 7 - 8 - 5 - 1 -.","Figure 1: Learning robust joint representations via multimodal
cyclic translations. Top: cyclic translations from a source modality
(language) to a target modality (visual). Bottom: the representation
learned between language and vision are further translated into the
acoustic modality, forming the final joint representation. In both
cases, the joint representation is then used for sentiment prediction.",0.75,0.3394,0.5447,423,321,1.318,638ef7a23d3aa76141ec2785d75671d1,images/2018/arxiv_0000506.png,https://arxiv.org/pdf/1812.07809v2.pdf
arxiv_0000507,Figure 507,scientific_figure,Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities,arXiv Research Authors,1812.07809v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities. arXiv:1812.07809v2,"Figure 2: MCTN architecture for two modalities: the source modal-
ity XS and the target modality XT . The joint representation ES⇆T
is obtained via a cyclic translation between XS and XT . Next, the
joint representation ES⇆T is used for sentiment prediction. The
model is trained end-to-end with a coupled translation-prediction
objective. At test time, only the source modality XS is required.",A diagram of a computer system with a.,"A detailed layout showing the different steps in the process of producing an encodent - rnn that shows key features, attributes, and data points in.","A comprehensive technical explanation of the networked and distributed rnn network for the region of the internet, including the network network, a virtual connection and a virtual networked networked with a virtual system, and a network.","Figure 2: MCTN architecture for two modalities: the source modal-
ity XS and the target modality XT . The joint representation ES⇆T
is obtained via a cyclic translation between XS and XT . Next, the
joint representation ES⇆T is used for sentiment prediction. The
model is trained end-to-end with a coupled translation-prediction
objective. At test time, only the source modality XS is required.",0.81,0.3324,0.5712,541,366,1.478,4d381e73ae5930084bcabd9aef364292,images/2018/arxiv_0000507.png,https://arxiv.org/pdf/1812.07809v2.pdf
arxiv_0000508,Figure 508,scientific_figure,Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities,arXiv Research Authors,1812.07809v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities. arXiv:1812.07809v2,"Figure 3: Hierarchical MCTN for three modalities: the source
modality XS and the target modalities XT1 and XT2. The joint
representation ES⇆T1 is obtained via a cyclic translation between
XS and XT1, then further translated into XT2. Next, the joint rep-
resentation of all three modalities, E(S⇆T1)→T2, is used for senti-
ment prediction. The model is trained end-to-end with a coupled
translation-prediction objective. At test time, only the source modal-
ity XS is required for prediction.",A diagram of the process for depicting.,"A detailed layout showing the process of decoderm and decoderptioning in an encoder that shows key features, attributes, and data points in detail.","A comprehensive technical explanation of the decoderm process for the encoder - rnn network, from the university of engineering and technology, london 2012 / 2011 / 2013 / 7 / 2011, part ii / 3 / 2 representing all details, specifications, and.","Figure 3: Hierarchical MCTN for three modalities: the source
modality XS and the target modalities XT1 and XT2. The joint
representation ES⇆T1 is obtained via a cyclic translation between
XS and XT1, then further translated into XT2. Next, the joint rep-
resentation of all three modalities, E(S⇆T1)→T2, is used for senti-
ment prediction. The model is trained end-to-end with a coupled
translation-prediction objective. At test time, only the source modal-
ity XS is required for prediction.",0.93,0.301,0.6155,561,407,1.378,854ade2475327e0c46add75be5490ede,images/2018/arxiv_0000508.png,https://arxiv.org/pdf/1812.07809v2.pdf
arxiv_0000509,Figure 509,scientific_figure,Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities,arXiv Research Authors,1812.07809v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities. arXiv:1812.07809v2,"Figure 4: Variations of our models: (a) MCTN Bimodal with cyclic translation, (b) Simple Bimodal without cyclic translation, (c) No-Cycle
Bimodal with different inputs of the same modality pair, and without cyclic translation, (d) Double Bimodal for two modalities without cyclic
translation, with two different inputs (of the same pair), (e) MCTN Trimodal with input from (a), (f) Simple Trimodal for three modalities,
with input as a joint representation taken from previous MCTN for two modalities from (b) or (c), (g) Double Trimodal with input from (d),
(h) Concat Trimodal which is similar to (b) but with input as the concatenation of 2 modalities, (i) Paired Trimodal using one encoder and 2
separate decoders for modality translations. Legend: black modality is ground truth, red (“hat”) modality represents translated output, blue
(“hat”) modality is target output from previous translation outputs, and yellow box denotes concatenation.",A diagram of the four main functions.,"A detailed layout showing the different types of electrical connections in a home system, including a thermosta.","A comprehensive technical explanation of the electrical control system for the home furnace and boiler, including thermosta, thermocol, and thermo, the controls, and more than thermometers and the same.","Figure 4: Variations of our models: (a) MCTN Bimodal with cyclic translation, (b) Simple Bimodal without cyclic translation, (c) No-Cycle
Bimodal with different inputs of the same modality pair, and without cyclic translation, (d) Double Bimodal for two modalities without cyclic
translation, with two different inputs (of the same pair), (e) MCTN Trimodal with input from (a), (f) Simple Trimodal for three modalities,
with input as a joint representation taken from previous MCTN for two modalities from (b) or (c), (g) Double Trimodal with input from (d),
(h) Concat Trimodal which is similar to (b) but with input as the concatenation of 2 modalities, (i) Paired Trimodal using one encoder and 2
separate decoders for modality translations. Legend: black modality is ground truth, red (“hat”) modality represents translated output, blue
(“hat”) modality is target output from previous translation outputs, and yellow box denotes concatenation.",0.75,0.2635,0.5068,1224,343,3.569,4ea48d53fc90a1d722c3db354ee4c00a,images/2018/arxiv_0000509.png,https://arxiv.org/pdf/1812.07809v2.pdf
arxiv_0000510,Figure 510,scientific_figure,Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities,arXiv Research Authors,1812.07809v2,cs,2018,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2018). Found in Translation: Learning Robust Joint Representations by Cyclic Translations Between Modalities. arXiv:1812.07809v2,"Figure 5: t-SNE visualization of the joint representations learned by
MCTN. Legend: red: videos with negative sentiment, blue: videos
with positive sentiment. Adding modalities and using cyclic transla-
tions improve discriminative performance and leads to increasingly
separable representations.",A diagram of different datas showing.,"A detailed layout showing the different characteristics of a human ' s memory model, from left to right.","A comprehensive technical explanation of the multiple - layer data visual for the different types of data visual in various languages and languages, from the beginning to end of the 20th century ago to end 20th century 1900 - 20th century century.","Figure 5: t-SNE visualization of the joint representations learned by
MCTN. Legend: red: videos with negative sentiment, blue: videos
with positive sentiment. Adding modalities and using cyclic transla-
tions improve discriminative performance and leads to increasingly
separable representations.",0.7814,0.2035,0.4924,561,166,3.38,7a333ce959c003a751f1204e97b68b2e,images/2018/arxiv_0000510.png,https://arxiv.org/pdf/1812.07809v2.pdf
arxiv_0000511,Figure 511,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 1. The logico-semantics part of Unsworth’S taxonomy [32] is shown, where blue
borders show extensions to Martinec and Salway [22] and underlined names were
changed by the authors, but have the same meaning.",A diagram of the process of a data.,A detailed layout showing the different components of a computer system and its components in action to make it work.,"A comprehensive technical explanation of a communication system for the user and user of a computer system, including the following functions and features of the system - based system - page 1 - in - text - file - inequire.","Fig. 1. The logico-semantics part of Unsworth’S taxonomy [32] is shown, where blue
borders show extensions to Martinec and Salway [22] and underlined names were
changed by the authors, but have the same meaning.",0.75,0.2548,0.5024,1175,867,1.355,44ac7ad611533b49578109e2b923faf0,images/2019/arxiv_0000511.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000512,Figure 512,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 1. The logico-semantics part of Unsworth’S taxonomy [32] is shown, where blue
borders show extensions to Martinec and Salway [22] and underlined names were
changed by the authors, but have the same meaning.",A diagram of a computer system with.,"A detailed layout showing the basic structure of a software application for a computer system, including a network diagram.","A comprehensive technical explanation of a computer system for the user of the device, including the control of the system and the application of the user ' s devices to the system, as well known information, the user, the system.","Fig. 1. The logico-semantics part of Unsworth’S taxonomy [32] is shown, where blue
borders show extensions to Martinec and Salway [22] and underlined names were
changed by the authors, but have the same meaning.",0.87,0.2523,0.5612,773,779,0.992,18428f8c546b43f025215125c0acb46c,images/2019/arxiv_0000512.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000513,Figure 513,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 2. Image-text classes sample and exemplify by Marsh and White, showing that
the authors use the concept of Abstractness Level to add more depth to their catego-
rizations.",A diagram of a tug boat on the water.,"A detailed layout showing the front of a boat on water with smoke coming out of it ' s chimney that shows key features, attributes, and data points.","A comprehensive technical explanation of the tug boat that is in the water with smoke coming out of it ' s back end and people on board watching it from the side of it and a boat in the front of the ship,.","Fig. 2. Image-text classes sample and exemplify by Marsh and White, showing that
the authors use the concept of Abstractness Level to add more depth to their catego-
rizations.",0.81,0.2808,0.5454,972,613,1.586,a98cb1a6013b50f622f4730b6166f603,images/2019/arxiv_0000513.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000514,Figure 514,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1",NULL,A diagram of a large boat with a red.,"A detailed layout showing the side of a large ship with a red and white hull, a blue and red bow and a flag on the stern.","A comprehensive technical explanation of a ship model for the 3d model industry, including the design and construction of the ships, and the components of its layouts, including engine, parts and components and equipment, and materials and design.",NULL,0.75,0.3539,0.5519,1002,647,1.549,31216c242d710789b982a4c6531c4009,images/2019/arxiv_0000514.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000515,Figure 515,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 2. Image-text classes sample and exemplify by Marsh and White, showing that
the authors use the concept of Abstractness Level to add more depth to their catego-
rizations.",A diagram of a boat in the water and.,"A detailed layout showing a tugboat and a smaller boat in a harbor with smoke coming from it that shows key features, attributes, and data points in.","A comprehensive technical explanation of a tugboat and its design for the future, including the hulls and hulls of it ' s hulls, and the sterns, as well as shown above the hull, an upper deck representing all details, specifications, and.","Fig. 2. Image-text classes sample and exemplify by Marsh and White, showing that
the authors use the concept of Abstractness Level to add more depth to their catego-
rizations.",0.81,0.2646,0.5373,595,277,2.148,0ce71a40fd43951c41c5d225d41d13d3,images/2019/arxiv_0000515.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000516,Figure 516,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1",Fig. 3. Overview of the encoder and classifier network.,A diagram of a network of different.,"A detailed layout showing the different components of a computer network and its components, including the processor, cpu, and other.","A comprehensive technical explanation of the computer network architecture of a computer network with multiple layers and different types of components, including a networked network, a network, and a network and a host, and an interlinking,.",Fig. 3. Overview of the encoder and classifier network.,0.81,0.263,0.5365,524,331,1.583,e81aaacc09fd7c659cd21e918a905973,images/2019/arxiv_0000516.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000517,Figure 517,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 4. Overview of the decoder network, whose input is the article embedding gener-
ated by the encoder (Figure 3).",A diagram of a typical computer.,"A detailed layout showing the different levels of an active internet network, including the basic architecture and the host layer.","A comprehensive technical explanation of the intel processor for the next generation of computers and tablets, including the core and mains, the processor and the main components, and its architecture, including a third level, ism,.","Fig. 4. Overview of the decoder network, whose input is the article embedding gener-
ated by the encoder (Figure 3).",0.93,0.2853,0.6077,595,351,1.695,be78332ca9ad959ef75341e1c0e68367,images/2019/arxiv_0000517.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000518,Figure 518,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 5. Three example results of the autoencoder network with the originals on the left
and the reproduced samples on the right. consists of 100 random samples for each of the three classes. We have evaluated
three different versions of the autoencoder and classifier networks. 1. CLscratch: Train the classifier network as well as the encoder network from
scratch, making it an end-to-end approach.
2. CLfreeze: Train the classifier network, but freeze the weights of the pre-
trained encoder network.
3. CLtransfer: Train the classifier network and finetune the pre-trained encoder
network at the same time.",A diagram of the plot shows that there.,A detailed layout showing the distribution of the temperature of a gas in the atmosphere of the earth from the sun.,A comprehensive technical explanation of the time and temperature curve for the titraon gass in the atmosphere of earth ' s atmosphere is shown below a line graphed by the h2 - to a data from the time - t.,"Fig. 5. Three example results of the autoencoder network with the originals on the left
and the reproduced samples on the right. consists of 100 random samples for each of the three classes. We have evaluated
three different versions of the autoencoder and classifier networks. 1. CLscratch: Train the classifier network as well as the encoder network from
scratch, making it an end-to-end approach.
2. CLfreeze: Train the classifier network, but freeze the weights of the pre-
trained encoder network.
3. CLtransfer: Train the classifier network and finetune the pre-trained encoder
network at the same time.",0.81,0.2715,0.5408,694,692,1.003,fbc7ee9d685c781612d3726b3c9a69b1,images/2019/arxiv_0000518.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000519,Figure 519,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1",NULL,A diagram of the temperature of a room.,"A detailed layout showing the image of a man walking down the street with a blue umbrella in hand that shows key features, attributes, and data.","A comprehensive technical explanation of the human brain and its functions in its environment, including the brain and the body of an animal that is connected to it ' s surroundings, by a cell line with a network of cells.",NULL,0.81,0.2312,0.5206,696,693,1.004,1388f1fb1a810312b01d0b93f82a84b9,images/2019/arxiv_0000519.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000520,Figure 520,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 5. Three example results of the autoencoder network with the originals on the left
and the reproduced samples on the right.",A diagram of the flow of water through.,A detailed layout showing the flow of data from the data center in different stages of development and development.,"A comprehensive technical explanation of the flow of fluid through a channel in a stream, using a plot and a diagram of flow through a stream in a channel through a flow through the stream, with a stream and a stream.","Fig. 5. Three example results of the autoencoder network with the originals on the left
and the reproduced samples on the right.",0.75,0.2336,0.4918,491,491,1.0,da96e152367cd1c5f4d6dd14541c5833,images/2019/arxiv_0000520.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000521,Figure 521,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 6. Examples of correctly as well as misclassified examples from our test set, along
with predicted and ground-truth labels.",A diagram of a plot showing the effect.,"A detailed layout showing the different frequencys of a wave and its amplitudes in a single line that shows key features, attributes, and data.","A comprehensive technical explanation of the fourier of the waveforms and the amplitude of the signal from the left, the waveforms are different, and the same, the same as well as shown with the wavefronters.","Fig. 6. Examples of correctly as well as misclassified examples from our test set, along
with predicted and ground-truth labels.",0.81,0.2845,0.5473,440,127,3.465,bc78075bd4b2d01e4c1e9d3d3132f14d,images/2019/arxiv_0000521.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000522,Figure 522,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1",NULL,A diagram of a square with three.,"A detailed layout showing the location of a large lake on the east coast of the united states, in red and green.","A comprehensive technical explanation of the great lakes from space, including nasa imagery and the earth ' s surface, including the great plains and the great basins, the great lake of the north and the lower plains, the east.",NULL,0.75,0.2487,0.4994,426,624,0.683,6d7c221b96128ba54a81b1c680d0ce3f,images/2019/arxiv_0000522.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000523,Figure 523,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1",NULL,A diagram of a machine that has a.,A detailed layout showing the two components of a microconver and the flow of water into each other.,"A comprehensive technical explanation of the process for an automated system, including a model and a diagram of how it works in a machine shop or factory or in a factory? 2 parts factory? - to be ready to be sold.",NULL,0.7071,0.2791,0.4931,1308,548,2.387,03f2f0c39595bf86de0debe64cf7a3be,images/2019/arxiv_0000523.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000524,Figure 524,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1",NULL,A diagram of a woman holding a laptop.,"A detailed layout showing the different angles of a computer screen, with a woman holding up a notebook.","A comprehensive technical explanation of a 3d - printed laptop screen for the classroom student ' s perspective, using three views of the computer screen and the students ' s face, with their hands, looking at the same image, the screen.",NULL,0.7214,0.271,0.4962,660,248,2.661,9693d3525ec5364f407b8e7521c1ca72,images/2019/arxiv_0000524.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000525,Figure 525,scientific_figure,"""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text",arXiv Research Authors,1901.07878v1,cs,2019,http://creativecommons.org/licenses/by/4.0/,"arXiv Research Authors (2019). ""Is this an example image?"" -- Predicting the Relative Abstractness Level of Image and Text. arXiv:1901.07878v1","Fig. 6. Examples of correctly as well as misclassified examples from our test set, along
with predicted and ground-truth labels.",A diagram of two images with different.,"A detailed layout showing two images of a wave cycle and a diagram of the waves that are moving that shows key features, attributes, and data points.","A comprehensive technical explanation of the mechanism of light - potential energy and the effect of the magnetic field on the surface of the earth ' s atmosphere is shown in figure 1, 5, 1, 4, 6, 3,.","Fig. 6. Examples of correctly as well as misclassified examples from our test set, along
with predicted and ground-truth labels.",0.87,0.2731,0.5716,663,444,1.493,e6bdb340fba82627b8e9857e1b2a9a5d,images/2019/arxiv_0000525.png,https://arxiv.org/pdf/1901.07878v1.pdf
arxiv_0000526,Figure 526,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 1: (a) Top images returned by QS-balanced for query “CEO” on Occupations dataset and
(b) top images returned by QS-balanced for query “smiling” on CelebA dataset. The first row
shows images returned by the algorithm using the diversity control matrix, the second row shows the
images with most similarity to the query, the third rows shows images with best combined scores,
i.e., minimum of
n
DSq
IF (I)
o I∈S for each IF .",A diagram of the heads of people in.,"A detailed layout showing a group of people in suits and ties with their names on them, including a man.","A comprehensive technical explanation of the top ten employees and their roles in the company ' s digital marketing strategy, including the names, images, and business figures of the employees, the top five members, and the top six,.","Figure 1: (a) Top images returned by QS-balanced for query “CEO” on Occupations dataset and
(b) top images returned by QS-balanced for query “smiling” on CelebA dataset. The first row
shows images returned by the algorithm using the diversity control matrix, the second row shows the
images with most similarity to the query, the third rows shows images with best combined scores,
i.e., minimum of
n
DSq
IF (I)
o I∈S for each IF .",0.7214,0.2131,0.4673,935,433,2.159,9170e7cca60313c337dfe7392eb8f5c5,images/2019/arxiv_0000526.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000527,Figure 527,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Top images returned by QS-balanced for query “CEO” on Occupations dataset and,A diagram of the faces of celebrities.,"A detailed layout showing the different expressions of people who are not related to each other, including their faces.","A comprehensive technical explanation of the many faces of the famous hollywood stars and their age and appearance in the movie, the bachelors, and what ' s the bachelor??? - tv shows here?? or the most.",Top images returned by QS-balanced for query “CEO” on Occupations dataset and,0.75,0.2733,0.5116,945,497,1.901,527d551c06c36f2060f39e8d447d50e6,images/2019/arxiv_0000527.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000528,Figure 528,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 1: (a) Top images returned by QS-balanced for query “CEO” on Occupations dataset and
(b) top images returned by QS-balanced for query “smiling” on CelebA dataset. The first row
shows images returned by the algorithm using the diversity control matrix, the second row shows the
images with most similarity to the query, the third rows shows images with best combined scores,
i.e., minimum of
n
DSq
IF (I)
o",A diagram of the top ten employees in.,"A detailed layout showing the many heads of people in different ages and races, from top to bottom that shows key features, attributes, and data.","A comprehensive technical explanation of the different company ' s employees and their roles in the company 's workflows and processes, including the company name, age, location, position, location and business name, and group,.","Figure 1: (a) Top images returned by QS-balanced for query “CEO” on Occupations dataset and
(b) top images returned by QS-balanced for query “smiling” on CelebA dataset. The first row
shows images returned by the algorithm using the diversity control matrix, the second row shows the
images with most similarity to the query, the third rows shows images with best combined scores,
i.e., minimum of
n
DSq
IF (I)
o",0.81,0.2887,0.5494,1224,432,2.833,0851329d4ac6b8e99da7870977f93232,images/2019/arxiv_0000528.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000529,Figure 529,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 2: A simple post-processing approach for ensuring diversity in image search. A small “diver-
sity control set” of images is taken as input, and (relevant) images are assigned a similarity score
with each image in the control set to create a diversity control matrix. These scores are combined
with the query scores provided in a black-box manner using an existing image search approach. A
summarization algorithm then selects the final images using this combined score. See Algorithm 1
for details. how well the image corresponds to the query. The smaller the score A(q, I), for a query q and image
I, the better the image corresponds to the query. Since our framework is meant to extend an existing
image retrieval model, we can assume that such a score can be efficiently computed for each query
and image pair. Image Similarity Score: Suppose that we also have a generic image similarity function sim,
which takes as input a pair of images, I1, I2, and calculates a score of similarity of the two images,
sim(I1, I2). For the sake of consistency, here again we will assume that the smaller the score, the
more similar are the images.
While the framework we propose is independent of the query matching algorithm or the image
similarity function, we will present a concrete example of such algorithms and functions in a later
section. We first see how we can use this score to rank our dataset. Diversity using a control set:
A ranking/summary with respect to the scores returned by A is
unlikely to be visibly diverse without further intervention in most cases, as shown by prior studies
[48]. To ensure visible diversity in the results, we use a diversity control set TF and a clustering
approach. The diversity control set TF is a small set of visibly diverse images and will be used to
enforce the diversity in the output; for example, if the summary is required to be gender-diverse,
then the diversity control set will have equal number of images of men and women.
For each control image IF ∈TF , using sim(·, ·) as the distance metric, we can learn the cluster
of images around IF , by sorting {sim(IF , I)}I∈S for each IF ∈TF . In other words, we can associate
each image I ∈S to an image in the control set to which I is most similar. Using diversity control sets with existing redundancy-reducing algorithms:
To ensure
we take into account both the query score from the black-box A and the diversity with respect to
the control set TF , we have to combine the scores A(q, ·) and sim(IF , ·). As mentioned earlier, a
popular approach to combining query similarity and diversity is to diversify across the entire feature",A diagram of the process of creating a.,"A detailed layout showing the process of processing data into a single - layer memory processor, including the processing process.","A comprehensive technical explanation of the process for the printing process, using the image basics and its components, as described by the following instructions in the diagram below link figure 1 / 3 / 5 / 2 / 4 / 3 -.","Figure 2: A simple post-processing approach for ensuring diversity in image search. A small “diver-
sity control set” of images is taken as input, and (relevant) images are assigned a similarity score
with each image in the control set to create a diversity control matrix. These scores are combined
with the query scores provided in a black-box manner using an existing image search approach. A
summarization algorithm then selects the final images using this combined score. See Algorithm 1
for details. how well the image corresponds to the query. The smaller the score A(q, I), for a query q and image
I, the better the image corresponds to the query. Since our framework is meant to extend an existing
image retrieval model, we can assume that such a score can be efficiently computed for each query
and image pair. Image Similarity Score: Suppose that we also have a generic image similarity function sim,
which takes as input a pair of images, I1, I2, and calculates a score of similarity of the two images,
sim(I1, I2). For the sake of consistency, here again we will assume that the smaller the score, the
more similar are the images.
While the framework we propose is independent of the query matching algorithm or the image
similarity function, we will present a concrete example of such algorithms and functions in a later
section. We first see how we can use this score to rank our dataset. Diversity using a control set:
A ranking/summary with respect to the scores returned by A is
unlikely to be visibly diverse without further intervention in most cases, as shown by prior studies
[48]. To ensure visible diversity in the results, we use a diversity control set TF and a clustering
approach. The diversity control set TF is a small set of visibly diverse images and will be used to
enforce the diversity in the output; for example, if the summary is required to be gender-diverse,
then the diversity control set will have equal number of images of men and women.
For each control image IF ∈TF , using sim(·, ·) as the distance metric, we can learn the cluster
of images around IF , by sorting {sim(IF , I)}I∈S for each IF ∈TF . In other words, we can associate
each image I ∈S to an image in the control set to which I is most similar. Using diversity control sets with existing redundancy-reducing algorithms:
To ensure
we take into account both the query score from the black-box A and the diversity with respect to
the control set TF , we have to combine the scores A(q, ·) and sim(IF , ·). As mentioned earlier, a
popular approach to combining query similarity and diversity is to diversify across the entire feature",0.81,0.3084,0.5592,761,301,2.528,987eac77fabab52eea3192213562bd59,images/2019/arxiv_0000529.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000530,Figure 530,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 2: A simple post-processing approach for ensuring diversity in image search. A small “diver-
sity control set” of images is taken as input, and (relevant) images are assigned a similarity score
with each image in the control set to create a diversity control matrix. These scores are combined
with the query scores provided in a black-box manner using an existing image search approach. A
summarization algorithm then selects the final images using this combined score. See Algorithm 1
for details.",A diagram of an information system.,A detailed layout showing the process of an automated data management system for the internet user and its uses.,"A comprehensive technical explanation of the process of the project using a data - driven system to create a 3d image and capture it to scale the process and view it from the object, including the other parts, in this diagram,.","Figure 2: A simple post-processing approach for ensuring diversity in image search. A small “diver-
sity control set” of images is taken as input, and (relevant) images are assigned a similarity score
with each image in the control set to create a diversity control matrix. These scores are combined
with the query scores provided in a black-box manner using an existing image search approach. A
summarization algorithm then selects the final images using this combined score. See Algorithm 1
for details.",0.75,0.2977,0.5239,1224,541,2.262,51ab8b24af913200f190e7c1bd5cdbd1,images/2019/arxiv_0000530.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000531,Figure 531,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 3: The plots show the fraction of images of women, dark-skinned people, and their intersec-
tion in the top 100 results of Google Image Search. (a) For gender, we also provide the comparison
with Google results from 2013 [48]. While the fraction of women in the top Google results seems
to have increased, the fraction of gender stereotypical images is still high ( 0.7 on an average). (b)
Majority of the top Google Images for every occupation corresponds to gender-stereotypical, fair-
skinned people, independent of the ground truth of percentage of Black people in the occupation.
For the rest of the minority groups, the fraction is partially dependent on the ground truth. 3.1.3
Intersection of Gender and Skin-tone 57% of the images have both a gender and skin-tone (binary) label. Amongst these, 27% of the
images are of fair-skinned men, 21% are of fair-skinned women, 6% are of dark-skinned men, and
3% are of dark-skinned women. Once again, the fraction of images of dark-skinned men and women
is relatively much smaller than the fraction of fair-skinned men and women, as seen from Figure 3b.
Furthermore, if we associate each occupation with its gender stereotype (for example, “Male” if the
fraction of men in the occupation is larger than the fraction and women, and “Female” otherwise),
then 35 out of 96 occupations do not have any images of dark-skinned gender anti-stereotypical
people in the top 100 results.
Figure 3b also provides an insight into the variation of fraction of images of different groups
(formed by intersection of gender and skin-tone) with respect to ground truth of fraction of Black
people in occupations. For almost all occupations, a large portion of the top 100 images are of
gender stereotypical fair-skinned people, further showing that current Google results for occupations
do correspond to the stereotypes. Interestingly, the fraction of images of gender stereotypical fair-
skinned people does not seem to be dependent on the ground truth. While this partition takes up
a significant portion of top 100 images, the fraction of images from other three minority partitions
seem to be partially dependent on the ground truth.
This lack of gender diversity in Google results from 2013 has also been explored in detail in the
paper by Kay et al. [48]; our updated dataset shows that the current Google results still suffer from
some of the gender diversity problems explored in that paper. Furthermore, our analysis also shows
the Google Image results are lacking in terms of skin-tone diversity and intersectional diversity.
We test the performance of QS-balanced and MMR-balanced algorithms on this Occupations
dataset and compare the results, in terms of diversity and accuracy, to top Google results.",A diagram of a line graph with a.,"A detailed layout showing a line graph of women and men in occupation around the world, with an orange dot at the bottom.","A comprehensive technical explanation of gender - specific men and women in occupation around the world, 1970 - 2010, with data from google / cosmoin com and the data source google / www / ly / tm / representing all details, specifications, and.","Figure 3: The plots show the fraction of images of women, dark-skinned people, and their intersec-
tion in the top 100 results of Google Image Search. (a) For gender, we also provide the comparison
with Google results from 2013 [48]. While the fraction of women in the top Google results seems
to have increased, the fraction of gender stereotypical images is still high ( 0.7 on an average). (b)
Majority of the top Google Images for every occupation corresponds to gender-stereotypical, fair-
skinned people, independent of the ground truth of percentage of Black people in the occupation.
For the rest of the minority groups, the fraction is partially dependent on the ground truth. 3.1.3
Intersection of Gender and Skin-tone 57% of the images have both a gender and skin-tone (binary) label. Amongst these, 27% of the
images are of fair-skinned men, 21% are of fair-skinned women, 6% are of dark-skinned men, and
3% are of dark-skinned women. Once again, the fraction of images of dark-skinned men and women
is relatively much smaller than the fraction of fair-skinned men and women, as seen from Figure 3b.
Furthermore, if we associate each occupation with its gender stereotype (for example, “Male” if the
fraction of men in the occupation is larger than the fraction and women, and “Female” otherwise),
then 35 out of 96 occupations do not have any images of dark-skinned gender anti-stereotypical
people in the top 100 results.
Figure 3b also provides an insight into the variation of fraction of images of different groups
(formed by intersection of gender and skin-tone) with respect to ground truth of fraction of Black
people in occupations. For almost all occupations, a large portion of the top 100 images are of
gender stereotypical fair-skinned people, further showing that current Google results for occupations
do correspond to the stereotypes. Interestingly, the fraction of images of gender stereotypical fair-
skinned people does not seem to be dependent on the ground truth. While this partition takes up
a significant portion of top 100 images, the fraction of images from other three minority partitions
seem to be partially dependent on the ground truth.
This lack of gender diversity in Google results from 2013 has also been explored in detail in the
paper by Kay et al. [48]; our updated dataset shows that the current Google results still suffer from
some of the gender diversity problems explored in that paper. Furthermore, our analysis also shows
the Google Image results are lacking in terms of skin-tone diversity and intersectional diversity.
We test the performance of QS-balanced and MMR-balanced algorithms on this Occupations
dataset and compare the results, in terms of diversity and accuracy, to top Google results.",0.81,0.3143,0.5622,702,511,1.374,86c838372de757bd9f4a7395b010b1f4,images/2019/arxiv_0000531.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000532,Figure 532,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Fraction of images of dark-skinned people and dark- skinned women in Google results Figure 3: The plots show the fraction of images of women, dark-skinned people, and their intersec- tion in the top 100 results of Google Image Search.",A diagram of black people in depicting.,A detailed layout showing the relationship of black people in occupations and ethnic growth in africa and asia and the extent of the black population.,"A comprehensive technical explanation of the black people in ocean ground truth - sea level plot, with the following data for the black person in the sea level of the data and the black line graphing data below the data data.","Fraction of images of dark-skinned people and dark- skinned women in Google results Figure 3: The plots show the fraction of images of women, dark-skinned people, and their intersec- tion in the top 100 results of Google Image Search.",0.75,0.3243,0.5372,700,525,1.333,a04c8f25247ba0f78b2f31856c79494d,images/2019/arxiv_0000532.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000533,Figure 533,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 3: The plots show the fraction of images of women, dark-skinned people, and their intersec-
tion in the top 100 results of Google Image Search. (a) For gender, we also provide the comparison
with Google results from 2013 [48]. While the fraction of women in the top Google results seems
to have increased, the fraction of gender stereotypical images is still high ( 0.7 on an average). (b)
Majority of the top Google Images for every occupation corresponds to gender-stereotypical, fair-
skinned people, independent of the ground truth of percentage of Black people in the occupation.
For the rest of the minority groups, the fraction is partially dependent on the ground truth.",A diagram of a scattery plot showing.,A detailed layout showing the relationship of two different types of data in a data visual system and data visual.,"A comprehensive technical explanation of the differences between the two types of data - driven data visualization and visual data visualisation in a data visualized data visual environment, with a / b / c / t / tm / t.","Figure 3: The plots show the fraction of images of women, dark-skinned people, and their intersec-
tion in the top 100 results of Google Image Search. (a) For gender, we also provide the comparison
with Google results from 2013 [48]. While the fraction of women in the top Google results seems
to have increased, the fraction of gender stereotypical images is still high ( 0.7 on an average). (b)
Majority of the top Google Images for every occupation corresponds to gender-stereotypical, fair-
skinned people, independent of the ground truth of percentage of Black people in the occupation.
For the rest of the minority groups, the fraction is partially dependent on the ground truth.",0.75,0.2656,0.5078,1224,570,2.147,4021b00064b568a2bb0c1591980be9f2,images/2019/arxiv_0000533.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000534,Figure 534,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 4: Occupations dataset: (a) Percentage of women in top 50 results vs ground truth of
percentage of women in occupations.
The images are generated using QS-balanced, MMR-
balanced and other baselines for the Occupations dataset. The figure shows that the image results
from QS-balanced and MMR-balanced, are more gender-balanced (see also Table 2) than image
results from other algorithms. While the fraction of images of women from QS-balanced is slightly
lower than MMR-balanced, the fraction of gender-anti-stereotypical images for both algorithms
is close (see Table 2). (b) Percentage of dark-skinned people in top 50 results vs ground truth of
percentage of Black people in occupations. The image results from QS-balanced are relatively
more balanced with respect to skin-tone; however, the fraction of images of dark-skinned people is
low for all algorithms.",A diagram of the relationship of women.,"A detailed layout showing the relationship of women in occupation and the number of men in the country, as well as the number.","A comprehensive technical explanation of the mean of women in occupation and gender in the united states, from the u s bureau for women ' s health research, 2003 - 2013 - 2013 source, and national survey, p df.","Figure 4: Occupations dataset: (a) Percentage of women in top 50 results vs ground truth of
percentage of women in occupations.
The images are generated using QS-balanced, MMR-
balanced and other baselines for the Occupations dataset. The figure shows that the image results
from QS-balanced and MMR-balanced, are more gender-balanced (see also Table 2) than image
results from other algorithms. While the fraction of images of women from QS-balanced is slightly
lower than MMR-balanced, the fraction of gender-anti-stereotypical images for both algorithms
is close (see Table 2). (b) Percentage of dark-skinned people in top 50 results vs ground truth of
percentage of Black people in occupations. The image results from QS-balanced are relatively
more balanced with respect to skin-tone; however, the fraction of images of dark-skinned people is
low for all algorithms.",0.75,0.2812,0.5156,717,533,1.345,99854739ae2ba2416c082d20acac924a,images/2019/arxiv_0000534.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000535,Figure 535,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Skin-tone diversity - Comparison with baselines Figure 4: Occupations dataset:,A diagram of the average likelihood of.,A detailed layout showing the distribution of black people in occupation from average male males in the area above.,"A comprehensive technical explanation of the black people in occupation and the black persons in occupation, by age and gender, as described by the data table below, in the data graph below, from the chart of the data.",Skin-tone diversity - Comparison with baselines Figure 4: Occupations dataset:,0.81,0.3163,0.5632,717,533,1.345,f17a0bbbcc66a34d0b9fedbbf2bfb88b,images/2019/arxiv_0000535.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000536,Figure 536,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 4: Occupations dataset: (a) Percentage of women in top 50 results vs ground truth of
percentage of women in occupations.
The images are generated using QS-balanced, MMR-
balanced and other baselines for the Occupations dataset. The figure shows that the image results
from QS-balanced and MMR-balanced, are more gender-balanced (see also Table 2) than image
results from other algorithms. While the fraction of images of women from QS-balanced is slightly
lower than MMR-balanced, the fraction of gender-anti-stereotypical images for both algorithms
is close (see Table 2). (b) Percentage of dark-skinned people in top 50 results vs ground truth of
percentage of Black people in occupations. The image results from QS-balanced are relatively
more balanced with respect to skin-tone; however, the fraction of images of dark-skinned people is
low for all algorithms.",A diagram of a line graph that shows.,A detailed layout showing the differences between the two graphs that are in the data visual system for each of the graphs.,"A comprehensive technical explanation of the data visual and data visual for a computer system, including a scattery plot, and a data visual of a sca data visual visual data visual representation of the visual data data visual,.","Figure 4: Occupations dataset: (a) Percentage of women in top 50 results vs ground truth of
percentage of women in occupations.
The images are generated using QS-balanced, MMR-
balanced and other baselines for the Occupations dataset. The figure shows that the image results
from QS-balanced and MMR-balanced, are more gender-balanced (see also Table 2) than image
results from other algorithms. While the fraction of images of women from QS-balanced is slightly
lower than MMR-balanced, the fraction of gender-anti-stereotypical images for both algorithms
is close (see Table 2). (b) Percentage of dark-skinned people in top 50 results vs ground truth of
percentage of Black people in occupations. The image results from QS-balanced are relatively
more balanced with respect to skin-tone; however, the fraction of images of dark-skinned people is
low for all algorithms.",0.81,0.2351,0.5226,1224,547,2.238,0635fc3118516b0aa7bbeb1c16133831,images/2019/arxiv_0000536.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000537,Figure 537,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 5: Occupations dataset: Diversity Control Sets used in the experiments.
The first two
diversity controls (a) and (b) are hand-picked while the last two (c) and (d) were randomly sampled
from the PPB dataset. D.3
Results for different diversity control sets As noted earlier, we use 4 different diversity control sets in our empirical evaluations. The results
presented in the paper correspond to evaluation using PPB-control set 1. We provide the diversity
comparison for different diversity control sets in Figure 6.",A diagram of a group of different.,A detailed layout showing the faces of different people and their names on a black background with multiple color variations.,"A comprehensive technical explanation of the various people involved in the project, including a woman and a man, and a group of men and women, all wearing different colored tie colors and suits and ties, smiling and hats, and ties.","Figure 5: Occupations dataset: Diversity Control Sets used in the experiments.
The first two
diversity controls (a) and (b) are hand-picked while the last two (c) and (d) were randomly sampled
from the PPB dataset. D.3
Results for different diversity control sets As noted earlier, we use 4 different diversity control sets in our empirical evaluations. The results
presented in the paper correspond to evaluation using PPB-control set 1. We provide the diversity
comparison for different diversity control sets in Figure 6.",0.75,0.2791,0.5146,1052,213,4.939,0e956f1c3b02d8c1b98180e2a295833d,images/2019/arxiv_0000537.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000538,Figure 538,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,and,A diagram of a group of people that.,"A detailed layout showing a group of people with different facial expressions and colors on their faces, including a man in a suit.","A comprehensive technical explanation of the lgbt - themed portraits of the participants of the event, including the men and women who were the same person on the photo, the team, in the photo credited and the other side,.",and,0.75,0.2617,0.5059,1054,228,4.623,df48b212fa88f00b79971da386170507,images/2019/arxiv_0000538.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000539,Figure 539,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 5: Occupations dataset: Diversity Control Sets used in the experiments.
The first two
diversity controls (a) and (b) are hand-picked while the last two (c) and (d) were randomly sampled
from the PPB dataset.",A diagram of a group of people and a.,A detailed layout showing the different people who are participating in the event and the names of each people.,"A comprehensive technical explanation of the many people who have been awarded for their work on the project, including the candidates and the winners in the competition and the competition at the top three categoriess for the event??? s.","Figure 5: Occupations dataset: Diversity Control Sets used in the experiments.
The first two
diversity controls (a) and (b) are hand-picked while the last two (c) and (d) were randomly sampled
from the PPB dataset.",0.7464,0.2887,0.5175,1224,866,1.413,1c463491c97deb767b43f3cc46d12076,images/2019/arxiv_0000539.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000540,Figure 540,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 6:
Occupations dataset:
Gender and skin-tone diversity comparison of results of QS-
balanced algorithm on different diversity control sets.
For gender, using any of the diversity
control sets results in a more gender-balanced output. For skin-tone, using PPB Control Set-1
results in the best results amongst all diversity control sets. For most occupations, the top Google
images are have much larger or much smaller fraction of images of dark-skinned people. D.4
Results for different compositions of diversity control sets To explicitly see the impact of diversity control on the diversity of the output of the algorithm,
we can vary the content of the diversity control set and observe the corresponding changes in the
results. We first vary the fraction of women in the diversity control set. The diversity control sets
are randomly chosen for the PPB-dataset, while maintaining the desired gender ratio. The results
for different diversity control sets are presented in Figure 7a. The figure shows that increasing the
fraction of women in the diversity control set leads to an increase in the fraction of women in the
output set.
Similarly, increasing the fraction of images of dark-skinned people in the diversity control set
leads to an increase in fraction of images of dark-skinned people in the output; this is shown in
Figure 7b. Finally, Figure 7c shows the impact of variation of images of dark-skinned women in the
control set on the output. While the fraction of dark-skinned women still increases, it seems to be
upper bounded by the fraction of images of dark-skinned women in the dataset.",A diagram of the correlation of women.,A detailed layout showing the different correlations for the women in occupation and gender in the overall age.,A comprehensive technical explanation of the relationship between men and women in occupational care and their role as a result of the study of women in occupation in the u s p9 - 2 % of their own positions of p9s.,"Figure 6:
Occupations dataset:
Gender and skin-tone diversity comparison of results of QS-
balanced algorithm on different diversity control sets.
For gender, using any of the diversity
control sets results in a more gender-balanced output. For skin-tone, using PPB Control Set-1
results in the best results amongst all diversity control sets. For most occupations, the top Google
images are have much larger or much smaller fraction of images of dark-skinned people. D.4
Results for different compositions of diversity control sets To explicitly see the impact of diversity control on the diversity of the output of the algorithm,
we can vary the content of the diversity control set and observe the corresponding changes in the
results. We first vary the fraction of women in the diversity control set. The diversity control sets
are randomly chosen for the PPB-dataset, while maintaining the desired gender ratio. The results
for different diversity control sets are presented in Figure 7a. The figure shows that increasing the
fraction of women in the diversity control set leads to an increase in the fraction of women in the
output set.
Similarly, increasing the fraction of images of dark-skinned people in the diversity control set
leads to an increase in fraction of images of dark-skinned people in the output; this is shown in
Figure 7b. Finally, Figure 7c shows the impact of variation of images of dark-skinned women in the
control set on the output. While the fraction of dark-skinned women still increases, it seems to be
upper bounded by the fraction of images of dark-skinned women in the dataset.",0.7464,0.3262,0.5363,726,562,1.292,b5887232aa08f8229ab61b8829f34e40,images/2019/arxiv_0000540.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000541,Figure 541,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Skin-tone diversity comparison Figure 6: Occupations dataset: Gender and skin-tone diversity comparison of results of QS- balanced algorithm on different diversity control sets. For gender, using any of the diversity control sets results in a more gender-balanced output. For skin-tone, using PPB Control Set-1 results in the best results amongst all diversity control sets. For most occupations, the top Google images are have much larger or much smaller fraction of images of dark-skinned people. D.4 Results for different compositions of diversity control sets To explicitly see the impact of diversity control on the diversity of the output of the algorithm, we can vary the content of the diversity control set and observe the corresponding changes in the results. We first vary the fraction of women in the diversity control set. The diversity control sets are randomly chosen for the PPB-dataset, while maintaining the desired gender ratio. The results for different diversity control sets are presented in Figure 7a. The figure shows that increasing the fraction of women in the diversity control set leads to an increase in the fraction of women in the output set. Similarly, increasing the fraction of images of dark-skinned people in the diversity control set leads to an increase in fraction of images of dark-skinned people in the output; this is shown in Figure 7b. Finally, Figure 7c shows the impact of variation of images of dark-skinned women in the control set on the output. While the fraction of dark-skinned women still increases, it seems to be upper bounded by the fraction of images of dark-skinned women in the dataset. 42",A diagram of black and red scatters.,A detailed layout showing the correlation of the rate of black in occupations in the u s and canada.,"A comprehensive technical explanation of the scatter plot for the 3 / 4 graph in which the black is the most important to the data and how??? - r = 1?? 2 /?? /? - representing all details, specifications, and configurations of the system components.","Skin-tone diversity comparison Figure 6: Occupations dataset: Gender and skin-tone diversity comparison of results of QS- balanced algorithm on different diversity control sets. For gender, using any of the diversity control sets results in a more gender-balanced output. For skin-tone, using PPB Control Set-1 results in the best results amongst all diversity control sets. For most occupations, the top Google images are have much larger or much smaller fraction of images of dark-skinned people. D.4 Results for different compositions of diversity control sets To explicitly see the impact of diversity control on the diversity of the output of the algorithm, we can vary the content of the diversity control set and observe the corresponding changes in the results. We first vary the fraction of women in the diversity control set. The diversity control sets are randomly chosen for the PPB-dataset, while maintaining the desired gender ratio. The results for different diversity control sets are presented in Figure 7a. The figure shows that increasing the fraction of women in the diversity control set leads to an increase in the fraction of women in the output set. Similarly, increasing the fraction of images of dark-skinned people in the diversity control set leads to an increase in fraction of images of dark-skinned people in the output; this is shown in Figure 7b. Finally, Figure 7c shows the impact of variation of images of dark-skinned women in the control set on the output. While the fraction of dark-skinned women still increases, it seems to be upper bounded by the fraction of images of dark-skinned women in the dataset. 42",0.7071,0.3529,0.53,717,562,1.276,9a58e555c411515d1c8c2bacff1bced4,images/2019/arxiv_0000541.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000542,Figure 542,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 6:
Occupations dataset:
Gender and skin-tone diversity comparison of results of QS-
balanced algorithm on different diversity control sets.
For gender, using any of the diversity
control sets results in a more gender-balanced output. For skin-tone, using PPB Control Set-1
results in the best results amongst all diversity control sets. For most occupations, the top Google
images are have much larger or much smaller fraction of images of dark-skinned people.",A diagram of a line graph showing the.,"A detailed layout showing the differences between the two different data visuals of a computer system, and a graphed data visual.","A comprehensive technical explanation of the differences between the two data visual models for the data visual model and the data model model, including the data models for each model of the model and their results, in the datas,.","Figure 6:
Occupations dataset:
Gender and skin-tone diversity comparison of results of QS-
balanced algorithm on different diversity control sets.
For gender, using any of the diversity
control sets results in a more gender-balanced output. For skin-tone, using PPB Control Set-1
results in the best results amongst all diversity control sets. For most occupations, the top Google
images are have much larger or much smaller fraction of images of dark-skinned people.",0.81,0.2385,0.5242,1224,564,2.17,e9935b04c1b3d1b9767442653eff1b94,images/2019/arxiv_0000542.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000543,Figure 543,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 7: Occupations dataset: Performance QS-balanced algorithm on diversity control sets with
different compositions. D.5
Results for different α values We vary the quality-fairness parameter α and look at its impact on the performance of our algo-
rithms. The diversity results are presented in Figure 8, while Figure 9 expands on the accuracy for
different alphas.",A diagram of a line graph with.,"A detailed layout showing a scattery plot of woman in occupation and men in women in occupation that shows key features, attributes, and data points.","A comprehensive technical explanation of the women in occupation and their likelihood for women in the u s a r c e, as described by a graph from the u n t s t n, for the women ' s, t v representing all details, specifications, and configurations of.","Figure 7: Occupations dataset: Performance QS-balanced algorithm on diversity control sets with
different compositions. D.5
Results for different α values We vary the quality-fairness parameter α and look at its impact on the performance of our algo-
rithms. The diversity results are presented in Figure 8, while Figure 9 expands on the accuracy for
different alphas.",0.87,0.3783,0.6241,717,575,1.247,988ddcb07d0d55b36e32f7f8f1fdc543,images/2019/arxiv_0000543.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000544,Figure 544,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Different fraction of dark-skinned in Diversity Control Set,A diagram of the black people in.,A detailed layout showing the differences between black and white people in occupation of their population in the united states.,"A comprehensive technical explanation of black people in occupation and the rise of black population in the united states, 1965 - 2010, by age of black americans and ethnicity, and ethnicity and race, 2000 - age, to - age.",Different fraction of dark-skinned in Diversity Control Set,0.75,0.3091,0.5295,731,555,1.317,fe26bd870d23b4d8a2c2d8ce5e44512f,images/2019/arxiv_0000544.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000545,Figure 545,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 7: Occupations dataset: Performance QS-balanced algorithm on diversity control sets with
different compositions.",A diagram of the different types of.,A detailed layout showing the differences between the three types of the data visualization model and the results of the model.,"A comprehensive technical explanation of the results of multiple data - driven datasetes in the study of the different types of datasets and types of the datasete, as described by the data - based dataseters.","Figure 7: Occupations dataset: Performance QS-balanced algorithm on diversity control sets with
different compositions.",0.87,0.262,0.566,1224,990,1.236,4d53c7d7358679cbe3160922af937021,images/2019/arxiv_0000545.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000546,Figure 546,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",A diagram of the average and average.,"A detailed layout showing the position of the v - value and the number of values in each type that shows key features, attributes, and data points.","A comprehensive technical explanation of the average and median - height of a single - cell phone signal, using the data source of the data library, and the data processing platform of the system, and application, as well as well.","Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",0.81,0.2773,0.5436,664,438,1.516,34b615cd221b2271c8d6c2d01cbe44aa,images/2019/arxiv_0000546.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000547,Figure 547,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 9: Occupations dataset: Accuracy of results of QS-balanced and MMR-balanced algo-
rithms for different α-values. For both QS-balanced and MMR-balanced, the fraction of gender anti-stereotypical images
increases as the α value increases. With an increase in fairness, a loss in accuracy is expected.
While the figure shows a small change in average query scores, the standard deviation of the scores
seem to be decreasing as well, showing that as α increases, the dependence on the query decreases.
Hence a balance between query similarity and diversity score has to be maintained by choosing an
appropriate value of α, such as 0.5. D.6
Results for different summary sizes While the results we have presented so have been with respect to a summary of size 50. However,
the size of the summary can depend on the application and the results in the first page of any web-
search application will depend on the size of the screen or the device being used. Correspondingly,
it is important to analyze the results for different summary sizes as well.",A diagram of the average time for a.,A detailed layout showing the data for the different stages of the signal in which the signal is displayed.,"A comprehensive technical explanation of the average and median - value of a sample of the data for a single - line graph shows the average data for each group of two - value groups of a given data in the data, p.","Figure 9: Occupations dataset: Accuracy of results of QS-balanced and MMR-balanced algo-
rithms for different α-values. For both QS-balanced and MMR-balanced, the fraction of gender anti-stereotypical images
increases as the α value increases. With an increase in fairness, a loss in accuracy is expected.
While the figure shows a small change in average query scores, the standard deviation of the scores
seem to be decreasing as well, showing that as α increases, the dependence on the query decreases.
Hence a balance between query similarity and diversity score has to be maintained by choosing an
appropriate value of α, such as 0.5. D.6
Results for different summary sizes While the results we have presented so have been with respect to a summary of size 50. However,
the size of the summary can depend on the application and the results in the first page of any web-
search application will depend on the size of the screen or the device being used. Correspondingly,
it is important to analyze the results for different summary sizes as well.",0.7321,0.2489,0.4905,664,438,1.516,380b4e0e21c579bbb377d274690da518,images/2019/arxiv_0000547.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000548,Figure 548,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",A diagram of a line graph with a line.,"A detailed layout showing the distribution of women in occupationss and men in other professionss, from the data.","A comprehensive technical explanation of the average age and age of women in occupationss, from a chart by the university of chicago, ny, 2003 - 2010 - 2011, 1, 2, 4 - 3, 52, 000 representing all details, specifications, and configurations of the.","Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",0.81,0.3297,0.5698,693,528,1.312,19a2c49c683ae7fe972f7134c55e81ec,images/2019/arxiv_0000548.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000549,Figure 549,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",A diagram of the average number of.,"A detailed layout showing the number of women in occupationss as per the average age of the woman that shows key features, attributes, and data.","A comprehensive technical explanation of the average mean of women in occupationss and age groups, based on a plot from the u s department of labor statistics department of statistics office, 2013 - u s / wm / n,.","Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",0.81,0.3387,0.5744,693,528,1.312,6a4f37820f3c70107fb75818aaeb9f8d,images/2019/arxiv_0000549.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000550,Figure 550,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",A diagram of a plot with different.,A detailed layout showing the various types of signal signals that can be used to determine the signal patterns.,"A comprehensive technical explanation of the mean and average mean of a single - variable variable variable - variable, or similar variable variable, variable variable or variable variableized, variableized variable variableised variable -.","Figure 8: Occupations dataset: Gender diversity and query similarity comparison of results of
QS-balanced and MMR-balanced algorithms for different α-values.",0.75,0.2272,0.4886,1224,508,2.409,8443a5a9f130ee70bbb658538e48d739,images/2019/arxiv_0000550.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000551,Figure 551,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 9: Occupations dataset: Accuracy of results of QS-balanced and MMR-balanced algo-
rithms for different α-values.",A diagram of the mean of time for two.,"A detailed layout showing the different results of each model, as well as the data in the chart below.","A comprehensive technical explanation of the current mean of the data visual for each of the following plots, from the data to the actual data visual data visual visual data source in the data source, from data source and data visual.","Figure 9: Occupations dataset: Accuracy of results of QS-balanced and MMR-balanced algo-
rithms for different α-values.",0.8943,0.2473,0.5708,1224,450,2.72,bfba0f86293d523452a1aa5b43f8059a,images/2019/arxiv_0000551.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000552,Figure 552,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 10: Occupations dataset: Variation of fraction of anti-stereotypical images vs size of summary
for all algorithms.",A diagram of the size and type of.,A detailed layout showing the size of summary for a single - sided sample of a single species of a species.,"A comprehensive technical explanation of the size and proportion of the measured area for the different types of materials that are required in the experiment for the first time of study, in the study, and the study is at the same time.","Figure 10: Occupations dataset: Variation of fraction of anti-stereotypical images vs size of summary
for all algorithms.",0.7321,0.2103,0.4712,712,533,1.336,a33fd1c779f01c592d93f98cdb95ebf8,images/2019/arxiv_0000552.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000553,Figure 553,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,NULL,A diagram of a line graph with.,A detailed layout showing the size of the different types of sunstanks compared to the average sunstank.,"A comprehensive technical explanation of the mean of average sunstanks in the u s and canada during the 1970s - 1990s, and the actual size of the same amount of summary of the total sunstanknakest sunstankest.",NULL,0.7214,0.2246,0.473,721,533,1.353,f88a55d8b3f886c3ddf3a424fc59bc6d,images/2019/arxiv_0000553.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000554,Figure 554,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 10: Occupations dataset: Variation of fraction of anti-stereotypical images vs size of summary
for all algorithms.",A diagram of the different types of.,"A detailed layout showing the data of several different types of graphs in the same area of a line that shows key features, attributes, and data.","A comprehensive technical explanation of the same time frame for each of these images, including the same size and shape as expected in the previous image, with the same frame for the same age and time frame, the same period.","Figure 10: Occupations dataset: Variation of fraction of anti-stereotypical images vs size of summary
for all algorithms.",0.87,0.2244,0.5472,1224,565,2.166,703086ddad66f1f02ca743fb11ed565f,images/2019/arxiv_0000554.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000555,Figure 555,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Figure 11: Occupations dataset: Fraction of anti-stereotypical images for summary size 10.,A diagram of the number of women who.,"A detailed layout showing the correlation of women in occupation and male in males in nursing care, and the average number of women.","A comprehensive technical explanation of the correlation of women in occupation and gender in the united states, based on the age of 26 years and 30 years old, 2009 - 2010 to 2012 - 2013 - 2017 - 2019 - 2018 -.",Figure 11: Occupations dataset: Fraction of anti-stereotypical images for summary size 10.,0.75,0.3362,0.5431,721,533,1.353,0ec04a84116a0918d0663881adeec45d,images/2019/arxiv_0000555.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000556,Figure 556,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Fraction of images of dark-skinned people vs ground truth Figure 11: Occupations dataset: Fraction of anti-stereotypical images for summary size 10. D.8 Occupation accuracy of QS-balanced algorithm Finally, we also present the accuracy of the results of QS-balanced algorithm. The accuracy is measured as the number of images in the summary belonging to the queried occupation. The results for this accuracy are presented in Figure 14. Note that accuracy is not a good measure of quality in this case; this is because a lot of occupations have similar looking images. For example, images of lawyers and financial analysts are very similar, images of doctors and pharmacists are very similar. Hence when using image similarity as a method of query matching, one cannot expect the matched images to always belong to the same query. This problem is relatively less visible for CelebA dataset, since in that case the query similarity algorithm is more specialized to the dataset. Figure 14: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for different occupations. For each occupation and its summary, we present the number of images belonging to that occupation in the summary, as well as the other occupation with highest number of images in the summary. 46",A diagram of a line graph with.,A detailed layout showing the total and total black people in occupation of males and females in the uk.,"A comprehensive technical explanation of the black people in occupation of african americans and their families, by age group, and age group of individuals, 1950 - 2010 - 2011 - 2013 - 2016 - 2014 - 2012 - 2012, p.","Fraction of images of dark-skinned people vs ground truth Figure 11: Occupations dataset: Fraction of anti-stereotypical images for summary size 10. D.8 Occupation accuracy of QS-balanced algorithm Finally, we also present the accuracy of the results of QS-balanced algorithm. The accuracy is measured as the number of images in the summary belonging to the queried occupation. The results for this accuracy are presented in Figure 14. Note that accuracy is not a good measure of quality in this case; this is because a lot of occupations have similar looking images. For example, images of lawyers and financial analysts are very similar, images of doctors and pharmacists are very similar. Hence when using image similarity as a method of query matching, one cannot expect the matched images to always belong to the same query. This problem is relatively less visible for CelebA dataset, since in that case the query similarity algorithm is more specialized to the dataset. Figure 14: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for different occupations. For each occupation and its summary, we present the number of images belonging to that occupation in the summary, as well as the other occupation with highest number of images in the summary. 46",0.7214,0.3271,0.5242,721,533,1.353,334c7b9c7bdea87c45593913462eeb7b,images/2019/arxiv_0000556.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000557,Figure 557,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 14: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations.
For each occupation and its summary, we present the number of images
belonging to that occupation in the summary, as well as the other occupation with highest number
of images in the summary.",A diagram of the number of people in.,"A detailed layout showing the number of people who have visited the country in the past year, from the u s.","A comprehensive technical explanation of the global trade and investment cycle in the world ' s largest markets, from 2010 to 2015, with the highest rates of trade and lowest gdp per capitas and the most countries in the u s.","Figure 14: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations.
For each occupation and its summary, we present the number of images
belonging to that occupation in the summary, as well as the other occupation with highest number
of images in the summary.",0.7321,0.1949,0.4635,3194,1250,2.555,1186ae7df9035da00ccedc31e675d2a9,images/2019/arxiv_0000557.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000558,Figure 558,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Figure 11: Occupations dataset: Fraction of anti-stereotypical images for summary size 10.,A diagram of a plot with a line graph.,A detailed layout showing the different types of data generated by each type of software and its application for the system.,A comprehensive technical explanation of the results of multiple test results for a single - cell phone system and other devices in a single cell phone network with a single device or two - cell device or three - cell system - mode connection.,Figure 11: Occupations dataset: Fraction of anti-stereotypical images for summary size 10.,0.75,0.224,0.487,1224,565,2.166,5a3c263c5bc976f5e6987f51cae0e8f2,images/2019/arxiv_0000558.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000559,Figure 559,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 14: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations.
For each occupation and its summary, we present the number of images
belonging to that occupation in the summary, as well as the other occupation with highest number
of images in the summary.",A diagram of a line graph shows the.,A detailed layout showing the number of different types of the same type of human activity in a single region.,"A comprehensive technical explanation of the u s - balanced algorithms in the u kfd system for the u n s - based system, the u a c e s - b and gl - n d systems representing all details, specifications, and configurations of the system components in.","Figure 14: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations.
For each occupation and its summary, we present the number of images
belonging to that occupation in the summary, as well as the other occupation with highest number
of images in the summary.",0.7429,0.1939,0.4684,1224,706,1.734,c4a112200346b4e925d45040f4cb4203,images/2019/arxiv_0000559.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000560,Figure 560,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Figure 12: Occupations dataset: Fraction of anti-stereotypical images for summary size 20.,A diagram of the relationship between.,"A detailed layout showing the distribution of women in occupation and the number of female workers in occupation, from the time.","A comprehensive technical explanation of the effects of women in occupation and age difference by race, race, age, and gender, and race age, from different races to age groups, 2000 - based countries, 2000 to 2010 to 2012.",Figure 12: Occupations dataset: Fraction of anti-stereotypical images for summary size 20.,0.81,0.3153,0.5627,721,533,1.353,ad028b6e9a63b1745b60256512d6a4e7,images/2019/arxiv_0000560.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000561,Figure 561,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Fraction of images of dark-skinned people vs ground truth Figure 12: Occupations dataset: Fraction of anti-stereotypical images for summary size 20. We also present the bar graph for when 1-norm is used, instead of cosine distance for similarity. In this case, the accuracy is much worse and this is reason for using cosine-distance over 1-norm distance for all our simulations. Figure 15: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for different occupations using 1-norm for similarity. E Additional Results on CelebA Dataset In this section, we present additional details and empirical results for CelebA dataset. The additional results correspond to varying different parameters in the algorithm, such as α value or the diversity control set. 47",A diagram of the data in this graph.,A detailed layout showing the various types of black people in occupation and the number of individuals who have been.,"A comprehensive technical explanation of the black people in occupation and the black person in occupation, based on the graph below, from the data sheet, 2010 - 2012 to 201107 / 10 / 12 / 3 / 7 / representing all details, specifications, and.","Fraction of images of dark-skinned people vs ground truth Figure 12: Occupations dataset: Fraction of anti-stereotypical images for summary size 20. We also present the bar graph for when 1-norm is used, instead of cosine distance for similarity. In this case, the accuracy is much worse and this is reason for using cosine-distance over 1-norm distance for all our simulations. Figure 15: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for different occupations using 1-norm for similarity. E Additional Results on CelebA Dataset In this section, we present additional details and empirical results for CelebA dataset. The additional results correspond to varying different parameters in the algorithm, such as α value or the diversity control set. 47",0.75,0.3132,0.5316,721,533,1.353,e48c18d7798222c4661c13eaee15ef9b,images/2019/arxiv_0000561.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000562,Figure 562,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 15: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations using 1-norm for similarity.",A diagram of the number of people who.,A detailed layout showing the number of people living in this small town by type of population and population.,"A comprehensive technical explanation of the global trade and investment cycle, based on the european market in 2013, with the u s and canada as the lowest u s, as of the us and canadas, p2, and canada.","Figure 15: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations using 1-norm for similarity.",0.7429,0.2021,0.4725,3194,1250,2.555,7fee93655d7f9433b3a81c4c3697efd2,images/2019/arxiv_0000562.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000563,Figure 563,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Figure 12: Occupations dataset: Fraction of anti-stereotypical images for summary size 20.,A diagram of a plot with several.,"A detailed layout showing various types of data and graphs in the same image, including the data and data.","A comprehensive technical explanation of the different types of mean and average mean for a single function of measurement to determine the number of samples per square metre9787, 8, 6, 0, 1, 000, 456.",Figure 12: Occupations dataset: Fraction of anti-stereotypical images for summary size 20.,0.7886,0.2507,0.5196,1224,565,2.166,8a5c7906c639cc33eec52b3233336179,images/2019/arxiv_0000563.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000564,Figure 564,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 15: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations using 1-norm for similarity.",A diagram of the average number of.,"A detailed layout showing the number of people who have reported an accident in a car accident, and how they.","A comprehensive technical explanation of the average air pressure for the entire planet, and the average surface temperature for each planet, is shown in this chart below tableaumental data from the following the data for the figure 1s.","Figure 15: Occupations dataset: Accuracy comparison of results of QS-balanced algorithm for
different occupations using 1-norm for similarity.",0.7393,0.2448,0.492,1224,525,2.331,f6fec5af5388c58c3a2f6f0e053474bf,images/2019/arxiv_0000564.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000565,Figure 565,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 13: Occupations dataset: (a) Comparison of accuracy, as measured using mean query similar-
ity scores, of top 50 results across all occupations. For each occupation, we also plot mean similarity
to the query control set and the standard deviation using the dotted lines. The mean similarity
score of results of all algorithms are close to each other, showing that using diversity control set
does not adversely impact the accuracy. (b) Comparison of non-redundancy scores. As expected,
the results from DET have the largest non-redundancy score, measured as the log of determinant
of the product of feature matrix the output images and its transpose. The non-redundancy scores of
QS-balanced and MMR-balanced are the lowest, perhaps due to enforcing fairness constraints
using the diversity control set.",A diagram of a plot with a line graph.,A detailed layout showing the average and marginal time of women in occupationss to the next generation of women.,"A comprehensive technical explanation of the mean of women in occupationss, as described by the data on this chart, is based on the data of the data generated by the actual data of each data and the data in the data.","Figure 13: Occupations dataset: (a) Comparison of accuracy, as measured using mean query similar-
ity scores, of top 50 results across all occupations. For each occupation, we also plot mean similarity
to the query control set and the standard deviation using the dotted lines. The mean similarity
score of results of all algorithms are close to each other, showing that using diversity control set
does not adversely impact the accuracy. (b) Comparison of non-redundancy scores. As expected,
the results from DET have the largest non-redundancy score, measured as the log of determinant
of the product of feature matrix the output images and its transpose. The non-redundancy scores of
QS-balanced and MMR-balanced are the lowest, perhaps due to enforcing fairness constraints
using the diversity control set.",0.75,0.3236,0.5368,689,529,1.302,c4b5fea3ce677fac0e581d49a21828ef,images/2019/arxiv_0000565.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000566,Figure 566,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Comparison of accuracy, as measured using mean query similar- ity scores, of top 50 results across all occupations. For each occupation, we also plot mean similarity to the query control set and the standard deviation using the dotted lines. The mean similarity score of results of all algorithms are close to each other, showing that using diversity control set does not adversely impact the accuracy.",A diagram of the number and age of.,"A detailed layout showing the number of women in occupation in various countries and their age ranges, including the age.","A comprehensive technical explanation of the mean for women in occupation, by age and gender, from the u s department of labor statistics office, 2003 - 2010 - 2011 - 2007, and 2013 - 2012 - 2013 - 2010,.","Comparison of accuracy, as measured using mean query similar- ity scores, of top 50 results across all occupations. For each occupation, we also plot mean similarity to the query control set and the standard deviation using the dotted lines. The mean similarity score of results of all algorithms are close to each other, showing that using diversity control set does not adversely impact the accuracy.",0.75,0.2764,0.5132,698,533,1.31,2aa9de538eaf889c07dee5733a77098a,images/2019/arxiv_0000566.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000567,Figure 567,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 13: Occupations dataset: (a) Comparison of accuracy, as measured using mean query similar-
ity scores, of top 50 results across all occupations. For each occupation, we also plot mean similarity
to the query control set and the standard deviation using the dotted lines. The mean similarity
score of results of all algorithms are close to each other, showing that using diversity control set
does not adversely impact the accuracy. (b) Comparison of non-redundancy scores. As expected,
the results from DET have the largest non-redundancy score, measured as the log of determinant
of the product of feature matrix the output images and its transpose. The non-redundancy scores of
QS-balanced and MMR-balanced are the lowest, perhaps due to enforcing fairness constraints
using the diversity control set.",A diagram of the two plots show.,"A detailed layout showing a number of data and information for the same region of the project, including the data.","A comprehensive technical explanation of the differences between the two data visual models in the data visual model and the data modeling model, from the data analysis model to the data model in the models in this set is also the data models.","Figure 13: Occupations dataset: (a) Comparison of accuracy, as measured using mean query similar-
ity scores, of top 50 results across all occupations. For each occupation, we also plot mean similarity
to the query control set and the standard deviation using the dotted lines. The mean similarity
score of results of all algorithms are close to each other, showing that using diversity control set
does not adversely impact the accuracy. (b) Comparison of non-redundancy scores. As expected,
the results from DET have the largest non-redundancy score, measured as the log of determinant
of the product of feature matrix the output images and its transpose. The non-redundancy scores of
QS-balanced and MMR-balanced are the lowest, perhaps due to enforcing fairness constraints
using the diversity control set.",0.75,0.2276,0.4888,1224,556,2.201,ceb8f513055a17fd6665b4992f1c6d36,images/2019/arxiv_0000567.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000568,Figure 568,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,(a) Diversity Control Set - 1,A diagram of a series of images of.,"A detailed layout showing multiple faces of people with different hair styles and colors, including a man ' s face.","A comprehensive technical explanation of a hair transplant and treatment procedure for men and women in the uk and europe, with photos of their faces on them, before and after being cut out and after the procedure, the procedure is taken.",(a) Diversity Control Set - 1,0.75,0.264,0.507,910,155,5.871,1466cc157e57f95cfe7168116f3c97fe,images/2019/arxiv_0000568.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000569,Figure 569,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,(b) Diversity Control Set - 2,A diagram of four images with people.,"A detailed layout showing multiple facial expressions of people with different hair styles and haircuts, including man, woman, and man.","A comprehensive technical explanation of the evolution of steve jobs in photoshopping, including many of them being self - created and not being a person ' s face ' s own ' s photo, ' s avatar, ' face.",(b) Diversity Control Set - 2,0.75,0.2524,0.5012,910,155,5.871,02868ec6efaa5ba77011996693d93959,images/2019/arxiv_0000569.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000570,Figure 570,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 16: CelebA dataset: Diversity Control Sets used in the for empirical evaluation on CelebA
dataset.",A diagram of a series of portraits of.,A detailed layout showing the faces of different people who are smiling and not looking at the camera in different ways.,"A comprehensive technical explanation of the many faces of the same person in a series of pictures, and what they look like each has a different face shape or appearance to them all of them??? with different faces??.","Figure 16: CelebA dataset: Diversity Control Sets used in the for empirical evaluation on CelebA
dataset.",0.75,0.2984,0.5242,921,498,1.849,09f4abd97faac493e3e1c1e6bb0474e1,images/2019/arxiv_0000570.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000571,Figure 571,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Diversity Control Set - 4 Figure 16: CelebA dataset: Diversity Control Sets used in the for empirical evaluation on CelebA dataset. 51,A diagram of a series of pictures of.,"A detailed layout showing a variety of people with different faces and hair styles, all in different colors.","A comprehensive technical explanation of the different faces of people in this picture, including men and women, and a woman with a flower on her head, and two men in hats and one man with a woman, and one.",Diversity Control Set - 4 Figure 16: CelebA dataset: Diversity Control Sets used in the for empirical evaluation on CelebA dataset. 51,0.7357,0.2991,0.5174,926,493,1.878,b2169f0ee613912f5cb2d2b7c4e206cd,images/2019/arxiv_0000571.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000572,Figure 572,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 16: CelebA dataset: Diversity Control Sets used in the for empirical evaluation on CelebA
dataset.",A diagram of the many faces of women.,"A detailed layout showing a series of different images of women with different faces and hair styles, including one woman ' s face.","A comprehensive technical explanation of the women ' s day at the womens conference in 2011 - 2012, presented by the department of civil affairs, is shown below a grid image of the image from the women, from the top row.","Figure 16: CelebA dataset: Diversity Control Sets used in the for empirical evaluation on CelebA
dataset.",0.75,0.338,0.544,1224,1341,0.913,9e73a1fa85a21c41443da68b8c7553ef,images/2019/arxiv_0000572.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000573,Figure 573,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 17: CelebA dataset: Gender and accuracy comparison of results of QS-balanced algorithm
for all queries.",A diagram of the correlation of women.,A detailed layout showing the same results of women with feature training set and the same data for each.,"A comprehensive technical explanation of the difference between women and men in training on the same level of training, with a line graph that shows how many people are using the same training conditions for each team member to be different levels.","Figure 17: CelebA dataset: Gender and accuracy comparison of results of QS-balanced algorithm
for all queries.",0.905,0.3029,0.6039,745,713,1.045,06df50ec8868e252b2aaff6a1d6e3bed,images/2019/arxiv_0000573.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000574,Figure 574,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Gender diversity comparison,A diagram of the correlation of women.,"A detailed layout showing the number of women with feature training sets in each class, based on the data.","A comprehensive technical explanation of the results of drug testing on men and women with feature training set 2 - 3, 5, 6, 8, 9, 9 years, or 10, or 12, or older or older representing all details, specifications, and configurations of the system.",Gender diversity comparison,0.8486,0.3144,0.5815,745,713,1.045,791f9d9de5cb68c9b2a652dcb991b968,images/2019/arxiv_0000574.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000575,Figure 575,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 18: CelebA dataset: Gender diversity and accuracy comparison of results of QS-balanced
algorithm on different diversity control sets. For all the control sets, the performance with respect
to gender diversity and accuracy seems to be similar.",A diagram of the differences in the.,A detailed layout showing the differences between diversity and diversity of women with attributes of age and gender.,A comprehensive technical explanation of the differences between women and men in the sport of tennis and tennis - - - tennis - pro - tennis com - usa net - pros com - pro com - women vs - twitter com - club.,"Figure 18: CelebA dataset: Gender diversity and accuracy comparison of results of QS-balanced
algorithm on different diversity control sets. For all the control sets, the performance with respect
to gender diversity and accuracy seems to be similar.",0.75,0.3237,0.5369,717,548,1.308,99ff2a5c39cfaea001bb2c7457287e06,images/2019/arxiv_0000575.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000576,Figure 576,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 18: CelebA dataset: Gender diversity and accuracy comparison of results of QS-balanced
algorithm on different diversity control sets. For all the control sets, the performance with respect
to gender diversity and accuracy seems to be similar.",A diagram of a line graph with.,A detailed layout showing the difference of women with attribue and diversity control for different sets of women.,"A comprehensive technical explanation of the average and marginal diversity of women in the united states and canada, and their total number of individuals in the country, as per 1, 2, 5 % of women with attribut.","Figure 18: CelebA dataset: Gender diversity and accuracy comparison of results of QS-balanced
algorithm on different diversity control sets. For all the control sets, the performance with respect
to gender diversity and accuracy seems to be similar.",0.75,0.3477,0.5489,696,548,1.27,53d988bd4287c4908af4eb35c2bccfda,images/2019/arxiv_0000576.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000577,Figure 577,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 17: CelebA dataset: Gender and accuracy comparison of results of QS-balanced algorithm
for all queries.",A diagram of two graphs with different.,A detailed layout showing the differences between a single - line and a dotted plot in a single line.,"A comprehensive technical explanation of the effect of the scatters on the data in the data visual file, and the actual data visual for the data source source, and image, and data, with the data, from the data.","Figure 17: CelebA dataset: Gender and accuracy comparison of results of QS-balanced algorithm
for all queries.",0.7707,0.2327,0.5017,1224,699,1.751,443926b40fd8172a05a18b3a315cb45d,images/2019/arxiv_0000577.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000578,Figure 578,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 18: CelebA dataset: Gender diversity and accuracy comparison of results of QS-balanced
algorithm on different diversity control sets. For all the control sets, the performance with respect
to gender diversity and accuracy seems to be similar.",A diagram of two plots showing.,A detailed layout showing the differences between the two data types and the data type of the data source.,"A comprehensive technical explanation of the difference between the two datas in the data visual field and the data driven data visual data visual fields of the data field of the field, as well as well described by the data, the data.","Figure 18: CelebA dataset: Gender diversity and accuracy comparison of results of QS-balanced
algorithm on different diversity control sets. For all the control sets, the performance with respect
to gender diversity and accuracy seems to be similar.",0.7286,0.239,0.4838,1224,543,2.254,bdcf59e70981e59d4fe2d9ef96501d48,images/2019/arxiv_0000578.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000579,Figure 579,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 19: CelebA dataset: Performance of QS-balanced algorithm on diversity control sets with
different compositions.",A diagram of a scattery plot showing.,A detailed layout showing the relationship of women with attributes and the number of women who have tested the method.,"A comprehensive technical explanation of women ' s average and average age differences in the age of women with attribute and age of men and women with adults, from the age, to age of older to age, the same women.","Figure 19: CelebA dataset: Performance of QS-balanced algorithm on diversity control sets with
different compositions.",0.75,0.3131,0.5315,717,567,1.265,a1b6e75319fb159755b70992d2b180cd,images/2019/arxiv_0000579.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000580,Figure 580,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Accuracy comparison Figure 19: CelebA dataset: Performance of QS-balanced algorithm on diversity control sets with different compositions. Figure 20: CelebA dataset: Non-redundancy comparison of our methods vs baselines. 53,A diagram of a line graph showing the.,"A detailed layout showing the difference of women with attribute and women with adotive that shows key features, attributes, and data points in.","A comprehensive technical explanation of the relationship between women and men in the age of 40 and 40, as described by the data in the table below, from the data released in the text attribuzzeq data table.",Accuracy comparison Figure 19: CelebA dataset: Performance of QS-balanced algorithm on diversity control sets with different compositions. Figure 20: CelebA dataset: Non-redundancy comparison of our methods vs baselines. 53,0.81,0.3305,0.5703,696,567,1.228,681c830c5c37f9359de6ea25271a76e7,images/2019/arxiv_0000580.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000581,Figure 581,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Figure 20: CelebA dataset: Non-redundancy comparison of our methods vs baselines.,A diagram of a line graph showing the.,"A detailed layout showing a data visual of the results of women with given feature in the study, as well as the mean of the result.","A comprehensive technical explanation of the results of the three women with glyfer feature for the three years of the study of the e - b - 3 - 2 - 3 0 - 1 - - 4 - 4 women with - 4 representing all details, specifications, and configurations of the.",Figure 20: CelebA dataset: Non-redundancy comparison of our methods vs baselines.,0.87,0.3303,0.6001,703,545,1.29,b44ffd3d0a695dc33a5b517792177c2b,images/2019/arxiv_0000581.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000582,Figure 582,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 19: CelebA dataset: Performance of QS-balanced algorithm on diversity control sets with
different compositions.",A diagram of two plots with different.,"A detailed layout showing the differences in the number of people who have visited the world, and how they are using social data.","A comprehensive technical explanation of the different types of the data in this picture are shown in a diagram and in a scatter diagram, each has a different type of a different function to find the same number of the same type.","Figure 19: CelebA dataset: Performance of QS-balanced algorithm on diversity control sets with
different compositions.",0.75,0.2338,0.4919,1224,673,1.819,696a4b1e77dc2465066b35ffcf88a791,images/2019/arxiv_0000582.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000583,Figure 583,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Figure 20: CelebA dataset: Non-redundancy comparison of our methods vs baselines.,A diagram of the relationship between.,A detailed layout showing the differences between the different mean and the mean of the mean and average mean.,"A comprehensive technical explanation of the difference between the two points in the plot and the same point on the line, which is the best fit for the data??? - based on the dotted line??, the line?.",Figure 20: CelebA dataset: Non-redundancy comparison of our methods vs baselines.,0.7464,0.2408,0.4936,1224,585,2.092,507f6f6c862162490e35f3cf22bbda60,images/2019/arxiv_0000583.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000584,Figure 584,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 21: CelebA dataset: Gender diversity and query similarity comparison of results of QS-
balanced and MMR-balanced algorithms for different α-values.",A diagram of a graph with the number.,"A detailed layout showing the number of different types of the data, with different data points and numbers.","A comprehensive technical explanation of the hypoplexe and the hygiemical effects of the temperature curves in the air, water cycle, and surface, with the temperature, and the temperature at different stages.","Figure 21: CelebA dataset: Gender diversity and query similarity comparison of results of QS-
balanced and MMR-balanced algorithms for different α-values.",0.7357,0.2756,0.5057,664,438,1.516,8c90e577766e506cde5424dde5e4f9c2,images/2019/arxiv_0000584.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000585,Figure 585,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,Fraction of images of gender anti-stereotypical people vs summary size,A diagram of the different types of.,A detailed layout showing the differences between the two types of the data visual and the data viewer for each data.,"A comprehensive technical explanation of the value distribution of a single - cell phone data system and its applications in the internet age, age, and status of the phone number of the cellphoners, from 1950 to today to today.",Fraction of images of gender anti-stereotypical people vs summary size,0.75,0.2657,0.5079,664,438,1.516,f9256765b61dde30ba445ccaaf6fc43c,images/2019/arxiv_0000585.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000586,Figure 586,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 22: CelebA dataset: Variation of fraction of gender anti-stereotypical images and accuracy
vs size of summary for all algorithms.",A diagram of the average and maximum.,"A detailed layout showing the size of summary of a model for a single - dimensional system, including a number of components.","A comprehensive technical explanation of the size of a computer monitor monitor, from the data visual library of the computer monitor company, inc, and its specmeta, 2005 - 2010 - 2007 - 2009 - 2010, n -.","Figure 22: CelebA dataset: Variation of fraction of gender anti-stereotypical images and accuracy
vs size of summary for all algorithms.",0.81,0.2133,0.5117,712,533,1.336,c681244cf1ce0118b14af78c9be3ee1a,images/2019/arxiv_0000586.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000587,Figure 587,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 22: CelebA dataset: Variation of fraction of gender anti-stereotypical images and accuracy
vs size of summary for all algorithms.",A diagram of a plot with a line graph.,"A detailed layout showing the size of the data and the amount of the time it takes to record that shows key features, attributes, and data points in.","A comprehensive technical explanation of the mean of the average and medianized data in the data visual library for the data retrieval system, cgd and rl - based on the same data source, c - based papers,.","Figure 22: CelebA dataset: Variation of fraction of gender anti-stereotypical images and accuracy
vs size of summary for all algorithms.",0.81,0.2243,0.5171,693,533,1.3,f4e80a9b116370fb1e0ad837d175f17c,images/2019/arxiv_0000587.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000588,Figure 588,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 21: CelebA dataset: Gender diversity and query similarity comparison of results of QS-
balanced and MMR-balanced algorithms for different α-values.",A diagram of a line graph with.,"A detailed layout showing the different types of data being generated by the same data size, and the different data.","A comprehensive technical explanation of the mean of a normal - sized plot for a single - digit data system, with a linearized plot for each data file and a data source, to be used as shown in the same time.","Figure 21: CelebA dataset: Gender diversity and query similarity comparison of results of QS-
balanced and MMR-balanced algorithms for different α-values.",0.75,0.2703,0.5101,1224,603,2.03,c644c880e95029da12ae77e8192fef54,images/2019/arxiv_0000588.png,https://arxiv.org/pdf/1901.10265v3.pdf
arxiv_0000589,Figure 589,scientific_figure,Implicit Diversity in Image Summarization,arXiv Research Authors,1901.10265v3,cs,2019,http://creativecommons.org/licenses/by/4.0/,arXiv Research Authors (2019). Implicit Diversity in Image Summarization. arXiv:1901.10265v3,"Figure 22: CelebA dataset: Variation of fraction of gender anti-stereotypical images and accuracy
vs size of summary for all algorithms.",A diagram of the different types of.,A detailed layout showing the various types of data collected from different sources to be used in a project.,"A comprehensive technical explanation of the various types of data in this chart and their corresponding datas are shown in the following slides below, and below the chart below the table 2 / 3 / 4 / 5 / 6 / 6.","Figure 22: CelebA dataset: Variation of fraction of gender anti-stereotypical images and accuracy
vs size of summary for all algorithms.",0.7393,0.2178,0.4785,1224,629,1.946,ec7d3179320e0b616feaa08958edeaa1,images/2019/arxiv_0000589.png,https://arxiv.org/pdf/1901.10265v3.pdf