id
int64
1
546k
paper_arxiv_id
stringlengths
10
13
path
stringlengths
1
249
caption
stringlengths
9
170k
label
stringlengths
6
192
name
float64
546,028
2305.01034v2
figs/imagenet_vary_delta.png
\caption{Varying spatial resolution \\ on ImageNet}
\label{fig:vary_delta}
null
546,000
2302.07253v2
figs/Fig2-arch-pos-hnmems.png
\caption{\textbf{Left}: Inside the ET block. The input token $\mathbf{x}^{(t)}$ passes through a sequence of operations and gets updated to produce the output token $\mathbf{x}^{(t+1)}$. The operations inside the ET block are carefully engineered so that the entire network has a global energy function, which decreases ...
\label{fig:ET_block}
null
546,001
2302.07253v2
figs/Fig3-many-recons-2.png
\caption{Reconstruction examples of our Energy Transformer using images from the ImageNet-1k validation set. \textit{Top row:} input images where 50\% of the patches are masked with the learned MASK token. \textit{Middle row}: output reconstructions after 12 time steps. \textit{Bottom row}: original images.}
\label{fig:rec-imgs}
null
546,002
2302.07253v2
figs/Fig4-InterpThroughTime.png
\caption{Token representations and gradients are visualized using the decoder at different times during the dynamics. The Energy Attention (ATTN) block contributes general structure information to the masked patches at {\color[HTML]{4FAFAF}{\em earlier}} time steps, whereas the Hopfield Network (HN) significantly sharp...
\label{fig:interp-dynamics}
null
546,003
2302.07253v2
figs/all_chn_mems.jpg
\caption{Visualizing a randomly selected 3025 patch memories of the 3072 learned by weight matrix in the Hopfield Network module (HN) of our model. These memories are vectors of the same dimensions $D$ as the patch tokens, stored as rows in the weight matrix $\mathbf{\xi}$. Each image patch is visualized using the mode...
\label{fig:all_chn_mems}
null
546,004
2302.07253v2
figs/kq_head_mems.jpg
\caption{Visualizing the token dimension of the ``key'' and ``query'' matrices of the attention as image patches. Each head is represented as a cell on the $4 \times 3$ grid above. We use the trained decoder of our model to visualize each weight. }
\label{fig:all_attn_mems}
null
546,005
2302.07253v2
figs/pos_bias_options.png
\caption{The cosine similarity between position biases of patches when the ET model is trained under different hyperparameter choices for $\beta$ (inverse temperature of the attention energy) and weight decay. Our ET sees a trend where smoother correlations are observed with smaller $\beta$ and weight decay.}
\label{fig:pos-bias-options}
null
546,006
2302.07253v2
figs/ET-Full-worst_recons.jpg
\caption{Reconstruction examples of images with the worst MSE from the IN1k validation set. \textit{Top row:} input images where 50\% of the patches are masked with the learned MASK token. \textit{Middle row}: all tokens reconstructed after 12 time steps. \textit{Bottom row}: original images.}
\label{fig:worst-imgs-model}
null
546,007
2302.07253v2
figs/ET-Full-best_recons.jpg
\caption{Reconstruction examples of images with the best (lowest) MSE from the IN1k validation set. \textit{Top row:} input images where 50\% of the patches are masked with the learned MASK token. \textit{Middle row}: all tokens reconstructed after 12 time steps. \textit{Bottom row}: original images.}
\label{fig:best-imgs-model}
null
546,008
2305.01034v2
figs/hypothesis_space_new.png
\caption{Dividing the hypothesis space of a task. % XXX Also IRF doesn't quite understand the multiple blue circles in part b; maybe in both a, b make the purple region much smaller to reflect the intuition (and later numbers!) that good generalization generally requires very large amounts of inductive bias? ...
\label{fig:hypothesis_space}
null
546,009
2305.01034v2
figs/all_diffs.png
\caption{Image Classification}
\label{fig:img_classification}
null
546,010
2305.01034v2
figs/Omniglot_diff.png
\caption{Few-shot Learning}
\label{fig:few_shot_learning}
null
546,011
2305.01034v2
figs/Mujoco_diffs.png
\caption{Reinforcement Learning}
\label{fig:rl}
null
546,012
2305.01034v2
figs/all_diffs_combined.png
\caption{Task difficulties of various benchmark tasks across domains.}
\label{fig:benchmarks}
null
546,013
2305.01034v2
figs/imagenet_vary_datasize.png
\caption{Varying training data \\ on ImageNet}
\label{fig:vary_train_data}
null
546,014
2305.01034v2
figs/reacher_vary_error.png
\caption{Varying desired error \\ on Reacher}
\label{fig:vary_error_rate}
null
546,015
2305.01034v2
figs/RL_diff.png
\caption{Varying intrinsic dim \\ on Cartpole}
\label{fig:vary_intrinsic_dim_rl}
null
546,016
2305.01034v2
figs/imagenet_vary_intrinsic_dim.png
\caption{Varying intrinsic dim \\ on ImageNet}
\label{fig:vary_intrinsic_dim_imagenet}
null
546,017
2305.01034v2
figs/imagenet_vary_datasize.png
\caption{Varying training data \\ on ImageNet}
\label{fig:vary_train_data}
null
546,018
2305.01034v2
figs/reacher_vary_error.png
\caption{Varying desired error \\ on Reacher}
\label{fig:vary_error_rate}
null
546,019
2305.01034v2
figs/RL_diff.png
\caption{Varying intrinsic dim \\ on Cartpole}
\label{fig:vary_intrinsic_dim_rl}
null
546,020
2305.01034v2
figs/imagenet_vary_intrinsic_dim.png
\caption{Varying intrinsic dim \\ on ImageNet}
\label{fig:vary_intrinsic_dim_imagenet}
null
546,021
2305.01034v2
figs/Wasserstein_trend_n.png
null
null
null
546,022
2305.01034v2
figs/Wasserstein_trend_m.png
null
null
null
546,023
2305.01034v2
figs/Wasserstein_trend_n.png
null
null
null
546,024
2305.01034v2
figs/Wasserstein_trend_m.png
null
null
null
546,025
2305.01034v2
figs/imagenet_vary_n_classes.png
\caption{Varying number of classes\\ on ImageNet}
\label{fig:vary_n_classes}
null
546,026
2305.01034v2
figs/imagenet_vary_delta.png
\caption{Varying spatial resolution \\ on ImageNet}
\label{fig:vary_delta}
null
546,029
2305.01034v2
figs/constructing_hypotheses.png
\caption{An example of constructing a hypothesis space for a binary-classification task with a 1 dimensional input. The black line indicates the data manifold and the blue and red dots on the line indicate training points. The hypothesis space is constructed using basis functions of three different frequencies; observe...
\label{fig:constructing_hypotheses}
null