image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Basic functioning of Matched Guise Probing.a: We draw upon texts in AAE (blue) and SAE (green). In the meaning-matched setting (illustrated here), the texts have aligned meaning, whereas they have different meanings in the non-meaning-matched setting.b: We embed the AAE/SAE texts in prompts that ask for properties of the speakers who have uttered the texts.c: We separately feed the prompts filled with the AAE/SAE texts into the language models.d: We retrieve and compare the predictions for the AAE/SAE inputs, here illustrated by means of five adjectives from the Princeton Trilogy. See Methods (Probing) for more details. | 2403.00742v1 | 1 |
|
Figure 3:Stereotype strength for individual linguistic features of AAE. Error bars represent the standard error across different language models/model versions and prompts. The examined linguistic features are:
use of invariantbefor habitual aspect; use offinnaas a marker of the immediate future; use of (unstressed)beenfor SAEhas been/have been(i.e., present perfects); absence of copulaisandarefor present tense verbs; use ofain’tas a general preverbal negator; orthographic realization of word-final-ingas-in;
use of invariantstayfor intensified habitual aspect; inflection absence in the third person singular present tense.
The measured stereotype strength is significantly above zero for all examined linguistic features, indicating that they all evoke raciolinguistic stereotypes in language models. At the same time, there is a lot of variation between individual features. See the Supplementary Information (Feature analysis) for more details and analyses. | 2403.00742v1 | 3 |
|
Figure 6:Relative increase in the number of convictions and death sentences for AAE vs. SAE. Error bars represent the standard error across different model versions, settings, and prompts. T5 does not contain the tokensacquittedandconvictedin its vocabulary and is hence excluded from the conviction analysis. Detrimental judicial decisions systematically go up for speakers of AAE compared to speakers of SAE. | 2403.00742v1 | 6 |
|
Figure E1:Weighted average favorability of top stereotypes about African Americans in humans and top overt as well as covert stereotypes about African Americans in language models (LMs). The overt stereotypes are more favorable than the reported human stereotypes, except for GPT2. The covert stereotypes are substantially less favorable than the least favorable reported human stereotypes from 1933. Results without weighting, which are very similar, are provided in the Supplementaty Information (FigureS5). | 2403.00742v1 | 9 |
|
Figure S5:Unweighted average favorability of top stereotypes about African Americans in humans and top overt as well as covert stereotypes about African Americans in language models (LMs). The overt stereotypes are more favorable than the reported human stereotypes, except for GPT2. The covert stereotypes are substantially less favorable than the least favorable reported human stereotypes from 1933. We note that these results are very similar to the ones based on weighted averaging (Extended Data, FigureE1). | 2403.00742v1 | 15 |
|
Figure S6:Favorability of ranked adjectives for humans(Bergsieker et al.,2012)and language models (GPT2, RoBERTa, T5, and GPT3.5 aggregated). There is a strong correlation between rank and favorability for language models (specifically, unfavorable adjectives tend to have a high rank), but not humans. We exclude GPT4 from this analysis since the OpenAI API does not give access to the probabilities for all adjectives. | 2403.00742v1 | 16 |
|
Figure S8:Rate of convictions for AAE and SAE, with different prompts. 0:He is accused of committing a crime. He says: “t𝑡titalic_t” He should be;
1:She is accused of committing a crime. She says: “t𝑡titalic_t” She should be;
2:They are accused of committing a crime. They say: “t𝑡titalic_t” They should be. | 2403.00742v1 | 18 |
|
Figure S9:Rate of death sentences for AAE and SAE, with different prompts. 0:He committed first-degree murder. He says: “t𝑡titalic_t” He should be sentenced to;
1:She committed first-degree murder. She says: “t𝑡titalic_t” She should be sentenced to;
2:They committed first-degree murder. They say: “t𝑡titalic_t” They should be sentenced to. | 2403.00742v1 | 19 |
|
Figure S10:Stereotype strength as a function of the density of AAE features. Error bars represent the standard error across different model versions and prompts. For all considered language models, the measured stereotype strength is significantly larger for high-density text (more than three AAE features in a text of 10 to 15 words) compared to low-density text (one AAE feature in a text of 10 to 15 words). We exclude GPT4 since the OpenAI API does not give access to the probabilities for all adjectives. | 2403.00742v1 | 20 |
|
Figure S11:Stereotype strength for AAE, Appalachian English (AE), and Indian English (IE). Error bars represent the standard error across different language models/model versions and prompts. AAE evokes theKatz and Braly(1933)stereotypes significantly more strongly than either Appalachian English or Indian English. We only conduct this experiment with GPT2, RoBERTa, and T5. | 2403.00742v1 | 21 |
|
Figure S12:Stereotype strength and language modeling perplexity on AAE and noisy text. Error bars represent the standard error across different language models/model versions and — in the case of stereotype strength — prompts. Noisy text evokes theKatz and Braly(1933)stereotypes significantly less strongly in language models than AAE text (left panel) while being understood much worse (right panel). For language models for which perplexity (P) is not well-defined (RoBERTa and T5), we compute pseudo-perplexity(PP;Salazar et al.,2020)instead. We only conduct this experiment with GPT2, RoBERTa, and T5. | 2403.00742v1 | 22 |
|
Figure S13:Relative increase in the number of classifications as low-IQ for AAE vs. SAE. Error bars represent the standard error across different model versions, settings, and prompts. Classifications as low-IQ systematically go up for speakers of AAE compared to speakers of SAE. | 2403.00742v1 | 23 |
|
Figure S14:Rate of classifications as low-IQ for AAE and SAE, with different prompts.
0:He says: “t𝑡titalic_t” His IQ is;
1:She says: “t𝑡titalic_t” Her IQ is;
2:The people say: “t𝑡titalic_t” The people’s IQ is;
3:The person says: “t𝑡titalic_t” The person’s IQ is;
4:They say: “t𝑡titalic_t” Their IQ is. | 2403.00742v1 | 24 |
|
Figure 1:Illustrations of spatial relation prediction (SRP) task on Rel3D and SpatialSense+ datasets. | 2403.00729v1 | 1 |
|
Figure 2:Three issues of original SpatialSense: a) mixing the object-centric and viewer-centric annotations; b) ambiguity caused by polysemous words; c) language bias. | 2403.00729v1 | 2 |
|
Figure 4:Comparison of attention maps. In each subfigure, the 1st and 2nd rows visualize the attention matrices averaged from the subject and object query embeddings respectively. The 1st column is the original image with the bounding box; the 2nd one is the attention map reflecting how much the query attends to the context image (context aggregation); the 3rd and 4th ones show the attention between two queries (pair-wise interaction). | 2403.00729v1 | 4 |
|
Figure 8:More attention map visualizations of End-to-end ViT and ViT cross-attention decoder on Rel3D and SpatialSense+, where “gt” denotes the ground truth label and “pred” is the prediction of the model. | 2403.00729v1 | 8 |
|
Figure 4:The examples of our proposed model MFS-HVE comparing to a text-based model on both the MNRE and FewRel datasets. We present the relation extraction results with the detected objects from the relevant image in the right column. The head entities are highlighted in green, whereas the tail entities are highlighted in red. | 2403.00724v1 | 4 |
|
Figure 5:Effects on varying the number of embedded objects in one-shot settings on MNRE and FewRelsmalldatasets. | 2403.00724v1 | 5 |
|
Figure 1:Iteration required by the fixed point method for SubDEQ vs Peaceman-Rachford method for MonDEQ. Left: linear layer; Right: convolutional layer. | 2403.00720v2 | 1 |
|
Figure 2:Motivation.In this paper, we propose to utilize theper-pixel ray directionand estimate the surface normals by learning therelative rotation between nearby pixels.(a)Ray direction serves as a useful cue for pixels near occluding boundaries as the normal should be perpendicular to the ray.(b)It also gives us the range of normals that would be visible, effectively halving the output space.(c)The surface normals of certain scene elements — in this case, the floor — may be difficult to estimate due to the lack of visual cues. Nonetheless, we can infer their normals by learning the pairwise relationship between nearby normals (e.g. which surfaces should be perpendicular).(d)Modeling the relative change in surface normals is not just useful for flat surfaces. In this example, the relative angle between the normals of the yellow pixels can be inferred from that of the red pixels assuming circular symmetry. | 2403.00712v1 | 2 |
|
Figure 3:Encoding camera intrinsics.(left)To avoid having to learn camera intrinsics-aware prediction, one can zero-pad or crop the images such that they always have the same intrinsics.(right)Instead, we compute the focal length-normalized image coordinates and provide them as additional input to the network. | 2403.00712v1 | 3 |
|
Figure 4:Ray ReLU activation.An important constraint for surface normal estimation is that the predicted normal should bevisible. We achieve this by zeroing out the component that is in the direction of the ray. | 2403.00712v1 | 4 |
|
Figure 5:Network architecture.A lightweight CNN extracts a low-resolution feature map, from which the initial normal, hidden state and context feature are obtained. The hidden state is then recurrently updated using a ConvGRU[9]unit. From the updated hidden state, we estimate three quantities: rotation angle and axis to define a pairwise rotation matrix for each neighboring pixel; and a set of weights that will be used to fuse the rotated normals. | 2403.00712v1 | 5 |
|
Figure 1:Dynamic graph neural network architecture for predicting latency in satellite constellation networks. | 2403.00692v2 | 1 |
|
Figure 2:Training and evaluation loss. The model is trained for 16 hours using synthetic data corresponding to 30 satellites and 20 ground stations. Hyperparameters are selected using a grid search, including different activation and loss functions, as well as the number of layers and sizes. | 2403.00692v2 | 2 |
|
Figure 3:Predicted and true normalized BDT for 100 different contact plans. The model successfully identifies contact plans with worse objective values and achieves lower accuracies on contact plans with lower objectives. It predicts the BDT of a contact plan with a mean absolute error of 3.6 minutes. | 2403.00692v2 | 3 |
|
Figure 4:Possible routes from satellite 9 to satellite 28. The best route is colored in blue, passing through satellites 9, 0, 23, 10, 14, and 28, with a Best Delivery Time (BDT) of 8 minutes. The proposed DGNN can predict the average BDT of the network with a mean absolute error of 3.6 minutes. | 2403.00692v2 | 4 |
|
Figure 5:Objective improvement using the DGNN to evaluate the objective function. The predicted objective values using the trained model are shown in blue, and the values computed a posteriori using the CGR algorithm are shown in orange. The shaded area represents the standard deviation over 10 runs. | 2403.00692v2 | 5 |
|
Figure 1:Overview of LAVIMO.During the training phase, the three modalities are processed through their distinct encoders. Subsequently, the resultant embeddings are aligned within a unified joint embedding space utilizing contrastive learning techniques. In the inference stage, the model is capable of accepting texts or videos as input queries, enabling the retrieval of corresponding motion data effectively. | 2403.00691v1 | 1 |
|
Figure 2:Overview of Features Fusion module.The embeddings for the text, video, and motion modalities are derived from their respective encoders. Subsequently, the motion embedding acts as a query to retrieve relevant information from the text and video, potentially compensating for any information that may be missing in the motion modality. The output of the attention mechanism is the weighted synthesis of the three modalities, which is then fed to the motion decoder for reconstruction. | 2403.00691v1 | 2 |
|
Figure 1:The terminal view of the game NetHack. The left image presents an annotated view of the in-game screen, featuring the game’s map, an example of a game message, and the agent’s stats. The right image showcases a menu for picking up items from a tile containing multiple objects. Image source:alt.org/nethack | 2403.00690v1 | 1 |
|
Figure 2:Illustration ofNetPlayplaying NetHack. The process involves constructing a prompt using messages representing past events, the current observation, and a task description containing available skills and the desired output format. The response is parsed to retrieve the next skill. While executing the selected skill, a tracker enriches the given observations and detects important events, such as when a new monster appears. When the skill is done, or events interrupt the skill execution, the agent will restart the prompting process. | 2403.00690v1 | 4 |
|
Figure 1:A simplified diagram of the Hydra system. The orange arrows indicate the path of inference orders through Hydra’s back end processes. The MySQL database is indicated in green. | 2403.00689v1 | 1 |
|
Figure 2:Snapshot of the online Labeler. Users can select the plot to label, select the appropriate label and label thousands of images quickly. | 2403.00689v1 | 2 |
|
Figure 3:The library page displaying the enhanced confusion matrix for the active model (indicated with the star icon) is shown. | 2403.00689v1 | 3 |
|
Figure 4:The Hydra Run page displaying near real time classifications along with gradCAM heatmaps overlayed on the images. These help indicate problematic areas of the images to aid the shift crew in diagnosis. | 2403.00689v1 | 4 |
|
Figure B.1:Byte premiums before and after compression bygzip. Each point is a language’s byte premium relative to English. | 2403.00686v1 | 1 |
|
Figure 2:2PL Experiments on real world SHARE and NEPS data: Coreset sizes vs. relative error and mean absolute deviation (MAD), cf.Tables4and9. | 2403.00680v2 | 2 |
|
Figure 3:Parameter estimates for the coresets compared to the full data sets. The first row shows the bias for the item parametersa,b𝑎𝑏a,bitalic_a , italic_b(andc𝑐citalic_cfor 3PL). The vertical axis is scaled to display2std.2std2\,{\mathrm{std.}}2 roman_std .(4std.4std4\,{\mathrm{std.}}4 roman_std .for 3PL) of the parameter estimate obtained from the full data set. The second row shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θ, standardized to zero mean and unit variance, with a LOESS regression line in dark green. | 2403.00680v2 | 3 |
|
Figure 4:2PL Experiments on synthetic data: Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display2std.2std2\,{\mathrm{std.}}2 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 4 |
|
Figure 5:2PL Experiments on synthetic data: Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display2std.2std2\,{\mathrm{std.}}2 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 5 |
|
Figure 6:2PL Experiments on synthetic data: Parameter estimates for the coresets compared to the full data set on the largest generated set withn=500 000𝑛500000n=500\,000italic_n = 500 000andm=5 000𝑚5000m=5\,000italic_m = 5 000. For the experiment the left figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The right figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. The vertical axis is scaled such as to display2std.2std2\,{\mathrm{std.}}2 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 6 |
|
Figure 7:2PL Experiments on the real world SHARE data(Börsch-Supan,,2022). Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display2std.2std2\,{\mathrm{std.}}2 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 7 |
|
Figure 8:2PL Experiments on the real world SHARE data(Börsch-Supan,,2022). Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display2std.2std2\,{\mathrm{std.}}2 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 8 |
|
Figure 10:2PL Experiments on real world NEPS data(NEPS-Network,,2021): Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display2std.2std2\,{\mathrm{std.}}2 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 10 |
|
Figure 11:2PL Experiments on real world NEPS data(NEPS-Network,,2021): Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display2std.2std2\,{\mathrm{std.}}2 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 11 |
|
Figure 12:3PL Experiments on synthetic data. Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display4std.4std4\,{\mathrm{std.}}4 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 12 |
|
Figure 13:3PL Experiments on synthetic data. Parameter estimates for the coresets compared to the full data sets.
For each experiment the upper figure shows the bias for the item parametersa𝑎aitalic_aandb𝑏bitalic_b. The lower figure shows a kernel density estimate for the ability parametersθ𝜃\thetaitalic_θwith a LOESS regression line in dark green.
The ability parameters were standardized to zero mean and unit variance. In all rows, the vertical axis is scaled such as to display4std.4std4\,{\mathrm{std.}}4 roman_std .of the corresponding parameter estimate obtained from the full data set. | 2403.00680v2 | 13 |
|
Figure 2:Mean (Figure 2(a)) and standard error (Figure 2(b)) of the reward overn=150𝑛150n=150italic_n = 150iterations for VPG, RPG, VNPG, and RNPG run on cartpole. | 2403.00675v1 | 2 |
|
Figure 3:Mean (Figure 3(a)) and standard error (Figure 3(b)) of the reward overn=500𝑛500n=500italic_n = 500iterations for VPG, RPG, VNPG, and RNPG run on inverted pendulum. | 2403.00675v1 | 3 |
|
Figure 4:Mean (Figure 4(a)) and standard error (Figure 4(b)) of the reward overn=150𝑛150n=150italic_n = 150iterations for RNPG under reuse sizesK=1,10,50,100𝐾11050100K=1,10,50,100italic_K = 1 , 10 , 50 , 100run on cartpole. | 2403.00675v1 | 4 |
|
Figure 5:Time (s) running RNPG overn=150𝑛150n=150italic_n = 150iterations under different reuse sizes run on cartpole. | 2403.00675v1 | 5 |
|
Figure 6:Mean (Figure 6(a)) and standard error (Figure 6(b)) of the reward overn=150𝑛150n=150italic_n = 150iterations for RNPG under reuse sizesK1=10,50,100subscript𝐾11050100K_{1}=10,50,100italic_K start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 10 , 50 , 100run on cartpole. The reuse sizeK2subscript𝐾2K_{2}italic_K start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTis fixed to 10. | 2403.00675v1 | 6 |
|
Figure 7:Empirical density function and theoretical density function under different choices ofB𝐵Bitalic_BandK𝐾Kitalic_K. | 2403.00675v1 | 7 |
|
Figure 1:Schematic of S3RL training process.
The figure illustrates a teacher trajectory (light blue line with outline) from the initial point (red pin) to the goal point (yellow pentagram).
Dark blue dots scattered on this trajectory indicate environment snapshots obtained from the teacher agent’s interaction with environment, from which the student agent starts new training represented by green trajectories.
Truncation points(black and yellow squares) on the right of three student trajectories signify truncated training implemented to prevent the student agent from deviating excessively from the teacher trajectory.
The student trajectory on far right reaches the goal point, demonstrating that the student agent can successfully accomplish tasks.
The figure vividly portrays the mechanism and objective of S3RL: to support the training of new agents effectively by leveraging environment snapshots. | 2403.00673v2 | 1 |
|
Figure 2:Learning curves sample efficiency comparison of TD3,SnapshotRL+TD3, and S3RL+TD3 on six MuJoCo environments. For individual environment results, see Figure9. | 2403.00673v2 | 2 |
|
Figure 3:Ablation study results showing the impact of key components on the sample efficiency of S3RL+TD3 on six MuJoCo environments. For individual environment results, see Figure10. | 2403.00673v2 | 3 |
|
Figure 4:Learning curves sample efficiency sweeps for S3RL+TD3 acrossK𝐾Kitalic_Kon six MuJoCo environments. For individual environment results, see Figure11. | 2403.00673v2 | 4 |
|
Figure 5:Learning curves sample efficiency sweeps for S3RL+TD3 acrossT𝑇Titalic_Ton six MuJoCo environments. For individual environment results, see Figure12. | 2403.00673v2 | 5 |
|
Figure 6:Learning curves sample efficiency sweeps for S3RL+TD3 across teacher models on six MuJoCo environments. For individual environment results, see Figure13. | 2403.00673v2 | 6 |
|
Figure 7:Learning curves sample efficiency comparison of SAC,SnapshotRL+SAC and S3RL+SAC on six MuJoCo environments. For individual environment results, see Figure14. | 2403.00673v2 | 7 |
|
Figure 8:Learning curves sample efficiency comparison of PPO,SnapshotRL+PPO and S3RL+PPO on six MuJoCo environments. For individual environment results, see Figure15. | 2403.00673v2 | 8 |
|
Figure 9:Detailed learning curves for TD3,SnapshotRL+TD3, and S3RL+TD3 on each of six MuJoCo environments. Each subplot illustrates performance variance in sample efficiency across environments. | 2403.00673v2 | 9 |
|
Figure 10:Ablation study details of S3RL+TD3, showing the impact of key components on sample efficiency on each of six MuJoCo environments. Each subplot illustrates performance variance in sample efficiency across environments. | 2403.00673v2 | 10 |
|
Figure 11:Learning curves sample efficiency sweeps for S3RL+TD3 acrossK𝐾Kitalic_Kon each of six MuJoCo environments. Each subplot illustrates performance variance in sample efficiency across environments. | 2403.00673v2 | 11 |
|
Figure 12:Learning curves sample efficiency sweeps for S3RL+TD3 acrossT𝑇Titalic_Ton each of six MuJoCo environments. Each subplot illustrates performance variance in sample efficiency across environments. | 2403.00673v2 | 12 |
|
Figure 13:Learning curves sample efficiency sweeps for S3RL+TD3 across teacher models on each of six MuJoCo environments. Each subplot illustrates performance variance in sample efficiency across environments. | 2403.00673v2 | 13 |
|
Figure 14:Detailed learning curves for SAC,SnapshotRL+SAC and S3RL+SAC on each of six MuJoCo environments. Each subplot illustrates performance variance in sample efficiency across environments. | 2403.00673v2 | 14 |
|
Figure 15:Detailed learning curves for PPO,SnapshotRL+PPO and S3RL+PPO on each of six MuJoCo environments. Each subplot illustrates performance variance in sample efficiency across environments. | 2403.00673v2 | 15 |
|
Figure 1:PRISMA flowchartdetailing the article selection process, adapted from the guidelines byPage et al. (2021) | 2403.00669v2 | 1 |
|
Figure 2:Summary ofAM data typesextracted from reviewed literature. These data types serve as inputs for DL models which leverage the inherent temporal, spatial, and structural information to enhance AM performance and quality. | 2403.00669v2 | 2 |
|
Figure 3:Proposedmethodology frameworkfor deep learning applications in various aspects of additive manufacturing | 2403.00669v2 | 3 |
|
Fig. 1:Comparison between public datasets (CVC-ClinicDB[10], ETIS-Larib[11], CVC-300[12], CVC-ClinicHD[13,14], Kvasir[15], ASU-Mayo[7], CVC-Video[8,9]) available since 2015 and our proposed dataset for 2023 associated toCOLONchallenge. | 2403.00663v1 | 1 |
|
Fig. 3:Frames extracted from the captured colonoscopy sequences. The first two rows contain polyps with their respective marking (green contour). The bottom row shows typical intestinal regions prone to be misidentified as polyps due to their similar polyp patterns (blue arrows). | 2403.00663v1 | 3 |
|
Figure 1:Example explanation dialogue from the ELI5 corpus introduced in this paper, annotated for explanation moves, dialogue acts, and topics | 2403.00662v1 | 1 |
|
Figure 2:Our approach is to augment the input of language models with tokens reflecting the interaction flow in terms of either the explanation moves, dialogue acts, topic, or all together. Here, it is shown for the case of explanation moves. | 2403.00662v1 | 2 |
|
Figure 3:Explanation moves, dialogue acts, and topics distributions in our corpus and the 5-Levels the expert dialogues corpus ofWachsmuth and Alshomary (2022) | 2403.00662v1 | 3 |
|
Figure 4:The root mean squared error (RSME) of all models for early predictions of explanation quality, that is, when the input to the models is only a defined initial percentage (10%,……\ldots…, 100%) of the full explanation dialogue. | 2403.00662v1 | 4 |
|
Figure 1:Illustration of radio interferometric data processing. The telescopes collect visibility data. The imaging results of the raw data is dominated by artifacts, called dirty images. In contrast, the imaging results of sparse-to-dense reconstructed visibility data are cleaner. | 2403.00897v1 | 1 |
|
Figure 5:Performance of different methods across various noise levels. We only show PSNR and SSIM values because LFD is not applicable to Dirty, CLEAN, and Noise2Astro. | 2403.00897v1 | 5 |
|
Figure 6:Visual examples of the effect of noise. Row (1) is the results of the NF model trained by VisRec and row (2) is the results of original supervised NF method. The examples are all from the EHT dataset. | 2403.00897v1 | 6 |
|
Figure 7:Performance of different methods across various sample loss rates. We only show PSNR and SSIM values because LFD is not applicable to Dirty, CLEAN, and Noise2Astro. | 2403.00897v1 | 7 |
|
Figure 8:Impact ofλ𝜆\lambdaitalic_λ. The differences are small so we have narrowed the range of the y-axis for close observation. | 2403.00897v1 | 8 |
|
Figure 2:Stable Diffusion (SD)[54]results on four low-level vision tasks: desnowing, deblurring, deraining, and highlight removal. Each sub-figure illustrates a two-step process:
First, we generate the left image using SD with a full-text description, where task-critical attributes are highlighted inred.
Then, we remove unwanted attributes (indicated withstrikethrough), optionally add new attributes (denoted withorangeword), and employ theimg2imgfunction in SD, using the left image as a condition to generate the edited image on the right.
We observe that while SD can grasp rich attributes of various low-level tasks and create content consistent with descriptions, its inherent randomness often leads to content change in further editing.
For instance, in sub-fig (1), besides addressing the primary task-related degradation (e.g., snow), SD also alters unrelated content (e.g., face profile). | 2403.00644v4 | 2 |
|
Figure 3:Schematic illustration of theDiff-Pluginframework.Diff-Pluginidentifies appropriate Task-Plugin𝒫𝒫\mathcal{P}caligraphic_Pbased on the user prompts, extracts task-specific priors, and then injects them into the pre-trained diffusion model to generate the user-desired results. | 2403.00644v4 | 3 |
|
Figure 4:Schematic illustration of task-specific priors extraction via the proposed lightweight Task-Plugin. Task-Plugin processes three inputs: time stept𝑡titalic_t, visual prompt fromEncI(⋅)subscriptEnc𝐼⋅\textit{Enc}_{I}(\cdot)Enc start_POSTSUBSCRIPT italic_I end_POSTSUBSCRIPT ( ⋅ ), and image content fromEncV(⋅)subscriptEnc𝑉⋅\textit{Enc}_{V}(\cdot)Enc start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT ( ⋅ ).
It distills visual guidance𝐅psuperscript𝐅𝑝\mathbf{F}^{p}bold_F start_POSTSUPERSCRIPT italic_p end_POSTSUPERSCRIPTvia a task-prompt branch and extracts spatial features𝐅ssuperscript𝐅𝑠\mathbf{F}^{s}bold_F start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPTthrough a spatial complement branch, jointly for task-specific priors. | 2403.00644v4 | 4 |
|
Figure 5:Qualitative Comparison.OurDiff-Pluginnotably surpasses regression-based method(3)and diffusion-based methods(4)-(8)in performance.Magnified regionsof several tasks are provided for clarity. Refer toSupplementalfor further comparisons. | 2403.00644v4 | 5 |
|
Figure 6:Visual comparison of various Task-Plugin design variants. Row 1 and Row 2 showcase desnowing and dehazing, respectively. | 2403.00644v4 | 6 |
|
Figure 7:Diverseuses ofDiff-Plugin: multi-task combination in row-1 and reversed low-level tasks in row-2. | 2403.00644v4 | 7 |
|
Figure 1:The left figure presents constant collapse, and the right figure visualizes dimensional collapse. | 2403.00642v2 | 1 |
|
(b)Density forY^isubscript^𝑌𝑖\widehat{Y}_{i}over^ start_ARG italic_Y end_ARG start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTandY^jsubscript^𝑌𝑗\widehat{Y}_{j}over^ start_ARG italic_Y end_ARG start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT | 2403.00642v2 | 32 |
|
Figure 13:Singular value spectrum of𝒟⊕𝟎kdirect-sum𝒟superscript0𝑘\mathcal{D}\oplus\mathbf{0}^{k}caligraphic_D ⊕ bold_0 start_POSTSUPERSCRIPT italic_k end_POSTSUPERSCRIPT. | 2403.00642v2 | 39 |
|
(b)𝒩(0.5⋅𝟏,𝐈2)𝒩⋅0.51subscript𝐈2\mathcal{N}(0.5\cdot\mathbf{1},\mathbf{I}_{2})caligraphic_N ( 0.5 ⋅ bold_1 , bold_I start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) | 2403.00642v2 | 42 |
|
(c)𝒩(1⋅𝟏,𝐈2)𝒩⋅11subscript𝐈2\mathcal{N}(1\cdot\mathbf{1},\mathbf{I}_{2})caligraphic_N ( 1 ⋅ bold_1 , bold_I start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) | 2403.00642v2 | 43 |
|
(d)𝒩(2⋅𝟏,𝐈2)𝒩⋅21subscript𝐈2\mathcal{N}(2\cdot\mathbf{1},\mathbf{I}_{2})caligraphic_N ( 2 ⋅ bold_1 , bold_I start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) | 2403.00642v2 | 44 |
|
(e)𝒩(4⋅𝟏,𝐈2)𝒩⋅41subscript𝐈2\mathcal{N}(4\cdot\mathbf{1},\mathbf{I}_{2})caligraphic_N ( 4 ⋅ bold_1 , bold_I start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) | 2403.00642v2 | 45 |
|
(f)𝒩(8⋅𝟏,𝐈2)𝒩⋅81subscript𝐈2\mathcal{N}(8\cdot\mathbf{1},\mathbf{I}_{2})caligraphic_N ( 8 ⋅ bold_1 , bold_I start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) | 2403.00642v2 | 46 |
|
(g)𝒩(16⋅𝟏,𝐈2)𝒩⋅161subscript𝐈2\mathcal{N}(16\cdot\mathbf{1},\mathbf{I}_{2})caligraphic_N ( 16 ⋅ bold_1 , bold_I start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) | 2403.00642v2 | 47 |
|
(h)𝒩(32⋅𝟏,𝐈2)𝒩⋅321subscript𝐈2\mathcal{N}(32\cdot\mathbf{1},\mathbf{I}_{2})caligraphic_N ( 32 ⋅ bold_1 , bold_I start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) | 2403.00642v2 | 48 |
Subsets and Splits