image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 7:Percentile 0.05 (a), median (b), and percentile 0.95 (c) of heights computed from canopy heights observed and predicted by our model in the 3,436 validation areas. Median of the points and 95% confidence interval computed by intervals of 5 m is given in blue. Each point represents one validation area. | 2501.10600v1 | 7 |
|
Figure 8:Examples of height changes due to logging activities for the Planet quad 0714-0954 for the periods 2022-03-01 to 2022-04-01 (a-c), 2022-04-01 to 2022-05-01 (d-f), and 2022-06-01 to 2022-07-01 (g-i). Each row of images corresponds to one of these periods and includes: a color composite image of the area (left), a color composite image of the subsequent month showing logging activity (middle), and a canopy height difference map (right). The image size is approximately 2.45 km×\times×2.45 km. | 2501.10600v1 | 8 |
|
Figure 9:Example of height time series in case of deforestation (points 1, 2, and 3), deforestation and regrowth (points 4, 5) and only regrowth (points 6). For each location, the value of the point and its 24 closest neighbors are presented. The locations of these points are displayed on our previously developed Planet NICFI-based deforestation product for the Planet quad 0681-0967 spanning from December 1, 2015, to February 1, 2024. The Planet NICFI quad covers approximately 19.5×\times×19.5 km. | 2501.10600v1 | 9 |
|
Figure 10:Example of artifacts in height prediction caused by haze, blur, clouds, and shade in the NICFI images before cloud correction for the Planet quad 0722-1038 basemap of 2022-03-01. Location 1 is a hazy but clear part of the image for reference, location 2 is a blurry part of the image, location 3 is a cloudy area with very high reflectance values, and location 4 is a shaded part of the image. Note that this image is classified as 100% clouds after filtering by our cloud model. The image size is∼similar-to\sim∼7.34×\times×7.34 km. | 2501.10600v1 | 10 |
|
Figure 11:Mean number of cloud-free observations for the∼similar-to\sim∼20 km×\times×20 km pixels corresponding to the Planet NICFI tiles. Only pixels with vegetation height greater than zero were considered in the mean computation. | 2501.10600v1 | 11 |
|
Figure 12:Map of the mean and percentile 97.5th of the Amazon forest tree canopy height on the period 2020-2024 (a-b). Each statistics is given for∼similar-to\sim∼20 km×\times×20 km pixels which correspond to a planet tile. Inside a tile, mean and percentile 97.5th are computed only with pixels that have≥\geq≥three non-clouded observation. | 2501.10600v1 | 12 |
|
Figure 13:Tree canopy height differences in meters between Tolan’s product (a), Pauls’s product (b), Lang’s product (c-d), and our canopy height product. For Lang’s product, we compared it to the mean of our product (c) and the 97.5th percentile (d), as Lang models the GEDI 98% relative height (RH98). Differences are shown for the∼similar-to\sim∼20 km×\times×20 km Planet NICFI tiles, considering only pixels with vegetation height greater than zero. | 2501.10600v1 | 13 |
|
Figure 15:Examples of giant trees identified by our Amazon canopy height model. For each individual tree (row), a Planet NICFI image is shown during its leafless period (first column), during its leaf-covered period (second column), and the third column presents our canopy height composite for the year 2020. The resolution of the NICFI image, preserved in this figure, is 4.78 m. | 2501.10600v1 | 15 |
|
Figure 2:NFENFE\mathrm{NFE}roman_NFEbetween the tensor of optimal VFs obtained via policy iteration low-rank PARAFAC decomposition in the grid-like setup. | 2501.10598v1 | 2 |
|
Figure 1:(a) Specific entropys𝑠sitalic_sand (b) pressurep𝑝pitalic_p(scaled byρT𝜌𝑇\rho Titalic_ρ italic_T) as a function of densityρ𝜌\rhoitalic_ρfor several isotherms. Solid line: the present work using flow matching method; dotted line: CMS19[11]; dashed line: REOS3[14]; dash-dotted line: the entropy reported by Moraleset al.[19]based on CEIMC simulations. Note the REOS3 entropy has been globallyshiftedto align with our results; see text for details. The inset of (b) presents the same pressure data by selected isochores, with density ranging from0.30.30.30.3(bottom) to0.7g/cm30.7gsuperscriptcm30.7\textrm{g}/\textrm{cm}^{3}0.7 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT(up). | 2501.10594v1 | 1 |
|
Figure 2:The line integral∮d(F/T)contour-integral𝑑𝐹𝑇\oint d(F/T)∮ italic_d ( italic_F / italic_T )(in unit of MJ/kg/K) computed along each closed local square loop of the REOS3 data. A schematic example of the loop is shown on the top-left corner. | 2501.10594v1 | 2 |
|
Figure 3:Several (a) specific entropy, (b) specific energy, (c) specific free energy, and (d) pressure (scaled byρT𝜌𝑇\rho Titalic_ρ italic_T) isotherms of our final EOS (solid lines), which consist of theab initiodata at high densities, the SCvH EOS at low densities, and an interpolation region between them, as indicated by the vertical lines. The dashed lines are the entropy and pressure isotherms of Miguelet al.[4], obtained by performing a single TI over the entire phase region of REOS3. Note we have added an extra entropy shift of−0.0057MJ/kg/K0.0057MJ/kg/K-0.0057\,\textrm{MJ/kg/K}- 0.0057 MJ/kg/Kto both the dashed lines and SCvH part of the solid lines in panel (a) to align with the experimental data at low pressures; see text for details. Panel (e) shows the local loop integral values∮d(F/T)contour-integral𝑑𝐹𝑇\oint d(F/T)∮ italic_d ( italic_F / italic_T )of our EOS, similar to Fig.2for the case of REOS3. | 2501.10594v1 | 3 |
|
Figure 4:Jupiter adiabats obtained from different EOSs. See texts for a detailed comparison between our result (red dashed line) and those derived from the REOS3 (black line)[22]and MH13 EOS (blue line)[12]. The yellow line shows the hydrogen-helium demixing curve for a protosolar mixture from Ref.[49]. | 2501.10594v1 | 4 |
|
Figure S1:Radial distribution functions of the two states atT0=3000Ksubscript𝑇03000KT_{0}=3000\textrm{K}italic_T start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 3000 K,ρ0=0.5g/cm3subscript𝜌00.5gsuperscriptcm3\rho_{0}=0.5\textrm{g}/\textrm{cm}^{3}italic_ρ start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 0.5 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPTandT1=3000Ksubscript𝑇13000KT_{1}=3000\textrm{K}italic_T start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 3000 K,ρ1=0.55g/cm3subscript𝜌10.55gsuperscriptcm3\rho_{1}=0.55\textrm{g}/\textrm{cm}^{3}italic_ρ start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 0.55 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT, respectively. The line labeled with “𝒙0subscript𝒙0\bm{x}_{0}bold_italic_x start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTflow” represents the new state obtained by transforming samples of state 0 using the velocity field trained with flow matching. | 2501.10594v1 | 5 |
|
Figure S2:Histograms for the distribution of forward and reverse work, defined in Eq. (S5), between the two states atT0=3000Ksubscript𝑇03000KT_{0}=3000\textrm{K}italic_T start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 3000 K,ρ0=0.5g/cm3subscript𝜌00.5gsuperscriptcm3\rho_{0}=0.5\textrm{g}/\textrm{cm}^{3}italic_ρ start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 0.5 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPTandT1=3000Ksubscript𝑇13000KT_{1}=3000\textrm{K}italic_T start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 3000 K,ρ1=0.55g/cm3subscript𝜌10.55gsuperscriptcm3\rho_{1}=0.55\textrm{g}/\textrm{cm}^{3}italic_ρ start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT = 0.55 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT. The vertical red (black) line is the estimated upper (lower) bound for free energy difference; see Eq. (S4) and the discussions therein. | 2501.10594v1 | 6 |
|
Figure S3:Same as Fig.S1, except the states 0 and 1 correspond to the uniform distribution and the reference state atT=5000K𝑇5000KT=5000\textrm{K}italic_T = 5000 K,ρ=1.4g/cm3𝜌1.4gsuperscriptcm3\rho=1.4\textrm{g}/\textrm{cm}^{3}italic_ρ = 1.4 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT, respectively. | 2501.10594v1 | 7 |
|
Figure S4:Same as Fig.S2, except the states 0 and 1 correspond to the uniform distribution and the reference state atT=5000K𝑇5000KT=5000\textrm{K}italic_T = 5000 K,ρ=1.4g/cm3𝜌1.4gsuperscriptcm3\rho=1.4\textrm{g}/\textrm{cm}^{3}italic_ρ = 1.4 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT, respectively. | 2501.10594v1 | 8 |
|
Figure S5:Several specific entropy isotherms of REOS3 computed by TI over a wide density range (0.01⩽ρ⩽5g/cm30.01𝜌5gsuperscriptcm30.01\leqslant\rho\leqslant 5\textrm{g}/\textrm{cm}^{3}0.01 ⩽ italic_ρ ⩽ 5 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT, main panel) and only theab initioregion at high densities (ρ⩾0.3g/cm3𝜌0.3gsuperscriptcm3\rho\geqslant 0.3\textrm{g}/\textrm{cm}^{3}italic_ρ ⩾ 0.3 g / cm start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPT, inset). The corresponding temperatures for the isotherms are (from bottom to top) 3000, 4000, 5000, 8000, 10000, 12000, and 15000K. The red (blue) solid lines correspond to the results of integrating along isotherm (isochore) first, as schematically illustrated on the top-right corner. For both regions, the reference point is set to the one with lowest temperature and density. The dashed lines are the entropy data produced by Miguelet al.[4]. | 2501.10594v1 | 9 |
|
Figure S6:Same as Fig.S5, except the reference point (for TI calculation of the solid lines) is set to be atT=15000K𝑇15000KT=15000\textrm{K}italic_T = 15000 Kand the lowest density. | 2501.10594v1 | 10 |
|
Figure S7:Several specific entropy isotherms of hydrogen at room conditions from the SCvH EOS[21](solid lines) and experimental data[46,47](dashed lines). The black markers correspond to the standard condition (1 bar, 298.15K). | 2501.10594v1 | 11 |
|
Figure S8:The hydrogen EOS (solid lines) constructed in the same way as Fig.3in the main text, except that theoriginal(but incorrect) absolute entropy of the SCvH EOS is used. The dashed lines are the original entropy and pressure isotherms of Miguelet al.[4], obtained by performing a single TI over the entire phase region of REOS3. | 2501.10594v1 | 12 |
|
Figure S9:Jupiter adiabats obtained from different EOSs. Compared to Fig.4of the main text, we also additionally show the adiabat derived from the CMS19 EOS[11], as well as the entropy of our EOS by colored regions. | 2501.10594v1 | 13 |
|
Figure 1:ColorGridvisualization demonstrating how assistant agent (follower) learns after observing human (leader) actions. | 2501.10593v1 | 1 |
|
Figure 3:Using a frozen expert leader trained with IPPO, we train a cold-started follower varying the training reward structure such that random block collection would have positive, neutral, and negative expected value (EV). Dotted lines are baselines of the expert leader, A* search leader, and A* search follower which routes to the last color picked up by the A* leader. We use seed 0 for these comparisons, except for the A* agent scores which are averaged across 100 seeds. | 2501.10593v1 | 3 |
|
Figure 1:Worker instructions for our conversational text collection task (weather communication situation). | 2501.10582v1 | 1 |
|
Figure 1.Rapid Occupational Training within the AI workforce structure that is illustrating different skill sets needed at each level. | 2501.10579v1 | 1 |
|
Figure 2.The training cycle of the AI technicians. The research instrumentation to measure the success of the training as well as to inform curricular development efforts is an integral part of the cycle. | 2501.10579v1 | 2 |
|
Figure 3.The evolution of the training curriculum reflects the evolution of the AI Technician role between 2020 and 2024. The original 16 weeks (one semester) long training has been extended to 32 weeks (2 semesters). | 2501.10579v1 | 3 |
|
Figure 4.Distribution of overall scores from the four iterations of the training. The content of the training was changing which makes the iterations not directly comparable. The decreasing variance in scores suggests improvements in the curriculum and teaching methods as well as better targeted selection of the trainees. | 2501.10579v1 | 4 |
|
Figure 5.Left: Pre/post knowledge check scores distribution over the last three iterations of the training. Right: Pre/post self-efficacy scores over the last three iterations of the training. | 2501.10579v1 | 5 |
|
Figure 1:AITK allows users to visualize the activations within a network. On the left, the network has correctly recognized the image as the digit 3. In the center, the network is unsure of whether the image represents the digit 3 or 5. On the right, the network has incorrectly classified the checkerboard image as the digit 4. | 2501.10576v1 | 1 |
|
Figure 3:AITK automatically generates graphs summarizing the network’s progress during training at reducing loss (at left) and improving accuracy (at right). | 2501.10576v1 | 3 |
|
Figure 1:The shuffling algorithm with an example.Top Panel: Algorithmic description of the shuffling procedure described in Section4. Bottom Panel: An example of the shuffling algorithm usingN=16𝑁16N=16italic_N = 16tokens. The first row (S=0𝑆0S=0italic_S = 0) corresponds to the unshuffled sequence. WhenS=1𝑆1S=1italic_S = 1, the tokens are split into41superscript414^{1}4 start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPTblocks first and then, the blocks are shuffled. The last rowS=2𝑆2S=2italic_S = 2shows the fully shuffled case where the tokens are randomly permuted. | 2501.10573v1 | 1 |
|
Figure 2:Average Cosine Similarity.Left Panel: average cosine similarity among tokens for a single prompt as a function of model layers. Right Panel: average cosine similarity averaged over2244224422442244prompts as a function of layers for the full shuffle (S=5𝑆5S=5italic_S = 5) and the structured case (S=0𝑆0S=0italic_S = 0). The color bar indicates the shuffle indexS𝑆Sitalic_S. The shaded regions indicate the standard deviation from the mean. All curves have been calculated for theLlamamodel. | 2501.10573v1 | 2 |
|
Figure 3:Intrinsic Dimension.Left Panel: intrinsic dimension for a single random prompt as a function of model layers. Right Panel: intrinsic dimension averaged over2244224422442244prompts as a function of layers for the full shuffle (S=5𝑆5S=5italic_S = 5) and the structured case (S=0𝑆0S=0italic_S = 0). The shaded regions indicate the standard deviation from the mean. The color bar indicates the shuffle indexS𝑆Sitalic_S. All curves have been calculated for theLlamamodel. | 2501.10573v1 | 3 |
|
Figure 4:Angle distribution between nearest neighbors.Left Panel: histogram of the angles between the first and second nearest neighbor at layer10101010of theLlamamodel for a single prompt for the full shuffle case and structured case. The dotted vertical lines indicate the average angle between the nearest neighbors in both cases. Right Panel: histogram of the average angle between the first and second nearest neighbor at layer10101010of theLlamamodel in the fully shuffled (orange) and structured case (blue). The histograms are computed from2244224422442244prompts in each case. | 2501.10573v1 | 4 |
|
Figure 5:Neighborhood Overlap.Left Panel: neighborhood overlap for a single random prompt as a function of model layers forkNN=2subscript𝑘NN2k_{\rm{NN}}=2italic_k start_POSTSUBSCRIPT roman_NN end_POSTSUBSCRIPT = 2. The colorbar indicates the shuffle indexS𝑆Sitalic_S. Right Panel: neighborhood overlap averaged over2244224422442244prompts as a function of layers for the full shuffle (S=5𝑆5S=5italic_S = 5) and the structured case (S=0𝑆0S=0italic_S = 0). The shaded regions indicate the standard deviation from the mean and the grey region indicates the region around the ID peak when the shuffled prompts have a lower NO than the structured prompts. All curves have been calculated for theLlamamodel. | 2501.10573v1 | 5 |
|
Figure 6:Correlation between intrinsic dimension and the average cross-entropy loss.Pearson coefficient between the logarithm of the intrinsic dimension and model loss for different models as a function of layers. The shaded regions indicate the standard deviation from the mean. The three curves correspond toLlama(orange),Mistral(magenta), andPythia(blue). Thep𝑝pitalic_p-values for the Pearson coefficients in this plot are below0.010.010.010.01except for the last layer inPythia. | 2501.10573v1 | 6 |
|
Figure 7:Correlating intrinsic dimension at the last layer to cross-entropy loss.The points in the following plots are calculated using the2244224422442244prompts considered in this paper for theLlamamodel. - (a) Left Panel: analysis of the correlation between the logits ID at scaling =2222(refer to Figure18in the Appendix for scaling =4,8484,84 , 8) and the contextual entropy to the average contextual entropy and (b) Right Panel: comparing the average contextual entropy to the average cross-entropy loss. | 2501.10573v1 | 7 |
|
Figure 9:ID profiles for shuffling for different models.(a) ID profiles for prompt number3218321832183218from Pile-10K for different models across different levels of shuffling. Lighter colors represent a higher shuffle index, and darker colors indicate a more structured prompt, and (b) ID profiles for Pile-10K prompt number, averaged over50505050prompts, for both structured and fully shuffled cases. Lighter colors indicate higher shuffle indices and darker colors represent a more structured prompt. The shaded regions show the standard deviation from the mean. | 2501.10573v1 | 9 |
|
Figure 10:Structured vs Shuffled ID for different range scalings.Intrinsic dimension at scaling =2,4,82482,4,82 , 4 , 8as a function of layer for the full shuffle and structured case for the average over all the prompts forLlama. | 2501.10573v1 | 10 |
|
Figure 12:Intrinsic Dimension.Left Panel: intrinsic dimension for a single prompt as a function of layers. Right Panel: intrinsic dimension averaged over2244224422442244prompts as a function of layers. The shaded regions indicate standard deviation from the mean. The curves correspond toLlama(orange),Mistral(magenta) andPythia(blue). | 2501.10573v1 | 12 |
|
Figure 13:Neighborhood Overlap.Left Panel: neighborhood overlap for a single prompt as a function of layers. Right Panel: intrinsic dimension averaged over2244224422442244prompts as a function of layers. Shaded regions indicate standard deviation from the mean. The curves correspond toLlama(orange),Mistral(magenta) andPythia(blue). | 2501.10573v1 | 13 |
|
(b)Unshuffled - GRIDE scale analysis for an unshuffled prompt (prompt number3218321832183218) across layers. | 2501.10573v1 | 16 |
|
Figure 15:Scale analysis for average ID profile.The ID profile averaged over2244224422442244prompts for range scaling =2,4,82482,4,82 , 4 , 8, with shaded regions indicating the standard deviation from the mean. | 2501.10573v1 | 18 |
|
Figure 16:Scale analysis for average NO profile.The neighborhood overlap profile averaged over2244224422442244prompts for range scaling =2,4,82482,4,82 , 4 , 8, with shaded regions indicating the standard deviation from the mean. | 2501.10573v1 | 19 |
|
Figure 17:Scale analysis for the correlation between intrinsic dimension and loss.Pearson coefficient between the logarithm of intrinsic dimension and model loss at scalings=2,4,8absent248=2,4,8= 2 , 4 , 8for different models. | 2501.10573v1 | 20 |
|
Figure 18:Scale analysis for the correlation between intrinsic dimension of logits and contextual entropy.Pearson coefficient between the logarithm of the intrinsic dimension of the logits and model contextual entropy for scalings=2,4,8absent248=2,4,8= 2 , 4 , 8forLlama. | 2501.10573v1 | 21 |
|
Figure 20:Summary of results for Opt-6.7B at the token-level.Left panel: The ID curve for Opt-6.7B for scaling =2,4,82482,4,82 , 4 , 8for prompt number3218321832183218from Pile-10K. We observe a peak around layer20202020as in the prompt level[20]. Middle panel: Spearman correlation between ID and loss for Opt-6.7B for different range scalings at the token level as a function of layers. Right panel: Scatter plot with theID𝐼𝐷IDitalic_I italic_D(y-axis) and the average cross-entropy loss (x-axis) at scaling =2222, layer17171717for the2244224422442244prompts we consider in this text. | 2501.10573v1 | 25 |
|
Figure 21:Intrinsic Dimension profile over training forPythia.Left Panel: intrinsic dimension profile for a single random prompt as a function of layers for different levels of training. Right Panel: intrinsic dimension averaged over50505050prompts as a function of layers for the untrained (orange) and trained (blue) model. The shaded regions indicate the standard deviation from the mean. | 2501.10573v1 | 26 |
|
Figure 1.Real-world applications often generate vast amounts of tabular data, making Data-Centric AI essential for optimizing performance. This survey explores the critical aspects of feature selection and generation, key to advancing tabular data-centric AI. | 2501.10555v1 | 1 |
|
Figure 3.Feature selection is framed as a RL problem, wherein the agent’s actions correspond to the selection of individual features. The optimization objective is to enhance performance on downstream tasks. | 2501.10555v1 | 3 |
|
Figure 4.Feature generation is formulated as a RL problem, where the agent’s actions involve selecting individual features and applying mathematical operations. The optimization goal is same as feature selection. | 2501.10555v1 | 4 |
|
Figure 5.RL extensively explores the feature space, generating abundant feature learning (feature selection/generation) knowledge. Generative AI captures this knowledge in a continuous embedding space, enabling gradient-based search to identify optimized feature spaces. | 2501.10555v1 | 5 |
|
Figure 1.HDC for image classification. HyperCam uses an HD classifier to perform face detection and identification tasks onboard low-power wireless camera platforms. | 2501.10547v1 | 1 |
|
Figure 3.Key operations of BSC.(1)Basis vectors are generated for every letter.(2)Binding of data creates a record.(3)Bundling of words creates a set.(4)Permutation is applied to create hypervectors on-the-fly. | 2501.10547v1 | 3 |
|
Figure 4.HyperCam overview. The HyperCam classifier runs onboard a low-power wireless camera platform and has three key components: image encoder, training algorithm, and inference algorithm. | 2501.10547v1 | 4 |
|
Figure 9.System Power Consumption. (1) camera plat- form in sleep mode (2) camera initialization, image capture, and inference, (3) data transmission, and (4) system returns to sleep mode. | 2501.10547v1 | 9 |
|
Figure 4:Embedding Table Partitioning strategies for TPU. There are three model parallelism strategies used: (a) table partitioning, (b) column partitioning, and (c) row partitioning. Each color coded block is an embedding table. The communication primitives used are AllToAll and AllGather. We employ a similar schematic[21], to illustrate dataflow. | 2501.10546v1 | 4 |
|
Figure 5:(a) Serialized and (b) Pipelined execution of TensorCore with SparseCore. Pipelining execution helps improve overlap, and thereby performance. | 2501.10546v1 | 5 |
|
Figure 6:Cost of the training service. The cost of input readers is much higher with LIG while the TPU cost is the same (assuming TPUs are not blocked by inputs). | 2501.10546v1 | 6 |
|
Figure 2:Schematic Diagram ofBRaIn:(A) Document Indexing & Retrieval, (B) Intelligent Relevance Feedback, (C) Query Expansion, and (D) Bug Localization | 2501.10542v1 | 2 |
|
Figure 2:Blow-up phenomenon. The observationsyi𝒙isubscript𝑦𝑖subscript𝒙𝑖y_{i}\boldsymbol{x}_{i}italic_y start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT bold_italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTare concentrated around the sphere𝝁+ρ−1/2Sp−1𝝁superscript𝜌12superscript𝑆𝑝1\boldsymbol{\mu}+\rho^{-1/2}S^{p-1}bold_italic_μ + italic_ρ start_POSTSUPERSCRIPT - 1 / 2 end_POSTSUPERSCRIPT italic_S start_POSTSUPERSCRIPT italic_p - 1 end_POSTSUPERSCRIPT. If⟨𝒘^,𝝁⟩‖𝒘^‖bold-^𝒘𝝁normbold-^𝒘\tfrac{\langle\boldsymbol{\hat{w}},\boldsymbol{\mu}\rangle}{\|\boldsymbol{\hat%
{w}}\|}divide start_ARG ⟨ overbold_^ start_ARG bold_italic_w end_ARG , bold_italic_μ ⟩ end_ARG start_ARG ∥ overbold_^ start_ARG bold_italic_w end_ARG ∥ end_ARGis big enough, a large proportion of the blue sphere lies in the shaded half-space. | 2501.10538v1 | 6 |
|
Figure 4:Illustration of𝒛⟂subscript𝒛perpendicular-to\boldsymbol{z}_{\perp}bold_italic_z start_POSTSUBSCRIPT ⟂ end_POSTSUBSCRIPTas a convex combination of𝒛¯isubscript¯𝒛𝑖\bar{\boldsymbol{z}}_{i}over¯ start_ARG bold_italic_z end_ARG start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT’s or, equivalently, as an orthogonal projection of the origin on the maximum margin hyperplane defined by them. | 2501.10538v1 | 8 |
|
Figure 1:Overall Architecture ofSCGT. The static functional connectivity features are used as input to the self-clustering graph attention blocks. Positional encodings are incorporated using Laplacian eigenvectors. The output node representations are concatenated to form the graph-level representation for prediction tasks. | 2501.16345v1 | 1 |
|
Figure 2:Visualization of the learnedθ1(l)subscriptsuperscript𝜃𝑙1\theta^{(l)}_{1}italic_θ start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTmatrix. Each row represents a brain region, and each column represents a functional community. Higher values indicate stronger associations. | 2501.16345v1 | 2 |
|
Figure 1:Examples of bivariate prediction regions with an80808080% coverage level for a toy example. | 2501.10533v2 | 1 |
|
Figure 2:Prediction regions for a bivariate unimodal dataset, conditional on a unidimensional input. The black, green, and yellow contours represent regions with nominal coverage levels of 20%, 40%, and 80%, respectively. | 2501.10533v2 | 2 |
|
Figure 4:Conditional coverage metrics across datasets sorted by size. CEC-X and CEC-Z should be minimized while WSC should approach1−α1𝛼1-\alpha1 - italic_α. | 2501.10533v2 | 4 |
|
Figure 9:Examples of prediction regions on a bivariate bimodal dataset, conditional on a unidimensional input. | 2501.10533v2 | 29 |
|
Figure 10:Panels 1 to 4: Trajectories of the log volume estimator with increasingK𝐾Kitalic_Kcompared to the true log volume (dashed line) for different output dimensionsd𝑑ditalic_d.
Panel 5: Log volume estimator withK=100𝐾100K=100italic_K = 100compared to the true log volume (dashed line). | 2501.10533v2 | 30 |
|
Figure 11:Marginal coverage and median region size with the base predictor MQF2across datasets sorted by size. | 2501.10533v2 | 31 |
|
Figure 13:Conditional coverage metrics with the base predictor Distributional Random Forests across datasets sorted by size. | 2501.10533v2 | 33 |
|
Figure 14:CD diagrams with the base predictor Distributional Random Forests based on 10 runs per dataset and method. | 2501.10533v2 | 34 |
|
Figure 16:CD diagrams based on Multivariate Gaussian Mixture Model parameterized by a hypernetwork withM=10𝑀10M=10italic_M = 10and 10 runs per dataset and method. | 2501.10533v2 | 36 |
|
Figure 17:Evolution of conditional coverage, marginal coverage and region sizes ofC-PCPas a function of the number of samplesK𝐾Kitalic_Kusing the base predictor MQF2. The metrics CEC-X𝑋Xitalic_X, and CEC-Z𝑍Zitalic_Zshould be minimized, while the marginal coverage and WSC should approach1−α1𝛼1-\alpha1 - italic_α(indicated by the dashed black line). The red line, obtained by linear regression, indicates the general trend. | 2501.10533v2 | 37 |
|
Figure 1:Architecture of theDFiN: The model processes input noise and noisy fingerprint through separate encoders forERBand complex features. After fusion (Eq.5), the features are decoded by respectiveERBandDFdecoders. The decodedERBfeatures are used to apply gains, while the noisy spectrum is filtered usingMFfilters. Combining gains with the filtered spectrum results in the enhanced spectrum estimateE𝐸Eitalic_E, which is inverted to producee(t)𝑒𝑡e(t)italic_e ( italic_t ). | 2501.10525v2 | 1 |
|
Figure 2:Stability analysis of theDFiNmodel on the VCTK-DEMAND dataset, showingΔΔ\Deltaroman_ΔSI-SDRas a function of the time difference (in seconds) between the noise mixture and noise fingerprints. | 2501.10525v2 | 2 |
|
Figure 3:Evaluation on the high-level categories of the classes in the ESC dataset. The radar chart illustrates theΔΔ\Deltaroman_ΔSI-SDRperformance boost across these high-level categories achieved by our model (DFiN) compared to the baseline (DFN). | 2501.10525v2 | 3 |
|
Figure 1:MBTA monthly bus ridership since January 2020, showing the current decline of 61.75% in ridership, compared to pre-pandemic levelsavi. | 2501.10514v1 | 1 |
|
Figure 2:Comprehensive IoT-enabled smart bus system: This diagram illustrates a public transit enhancement framework with three main components: data processing and dynamic scheduling (managed by a centralized cloud platform that processes real-time transit data and employs an AI-powered prediction model to adjust bus schedules dynamically), IoT-enabled buses (which continuously exchange data with the cloud to provide up-to-date information and improved service quality), and a mobile app for passengers to access the bus departure time information. | 2501.10514v1 | 2 |
|
Figure 3:An overview of the bus departure time prediction system: The system integrates three key data sources (transit operations data, meteorological data, and bus stop data). The preprocessing stage involves removing invalid entries and detecting outliers. Feature extraction generates both input and output features for the model. An FCNN with a one-bus stop lookback window is trained using the MSE loss function. In the prediction stage, the system provides real-time estimates of bus departure time deviations. | 2501.10514v1 | 3 |
|
Figure 6:Average departure time deviation across routes in preprocessed MBTA bus departure times 2023 dataset. | 2501.10514v1 | 6 |
|
Figure 7:Segmenting each trip into sequential segments of bus stops from the start point (SP) to the endpoint (EP), with each trip containing between 2 and 14 stops. | 2501.10514v1 | 7 |
|
Figure 8:The image illustrates the application of a bus departure time prediction model using an FCNN for two bus routes. It shows the scheduled and actual departure times, highlighting deviations with colored arrows. Red right-facing arrows indicate late departures, and green left-facing arrows indicate early departures. | 2501.10514v1 | 8 |
|
Figure 9:The model processes 173 input features, including time, distance, weather, and traffic data, through its hidden layers. The illustrated structure represents the model with optimal performance in terms of accuracy and computational complexity, with three hidden layers comprising 512, 128, and 64 neurons, respectively. The final output layer provides departure time deviation predictions. | 2501.10514v1 | 9 |
|
Figure 10:The performance of the proposed model across all bus routes in the preprocessed MBTA bus departure times 2023 test set. The average RMSE for all routes in the test set is 77.8312 s. Among the prediction errors, bus route 32 has the greatest RMSE, with a value of 240.61 s. | 2501.10514v1 | 10 |
|
Figure 1:Left:Failure cases of leading MLLMs, such as LLaVA-OV[47], Qwen2-VL[103], GPT-4o[30]and GeminiPro1.5[96], onbasicquestions related to face understanding.Right:Performance comparison of top models across the14141414tasks included in the benchmark. | 2501.10360v1 | 1 |
|
Figure 3:FaceXBench examplescover a total of 14 tasks, addressing various aspects of face understanding. Each question may consist of single or multiple images. Every question includes four options, with only one correct answer. The options are strategically designed to prompt the model to analyze carefully before selecting an option. | 2501.10360v1 | 5 |
|
Figure 4:(a) Performance of top-5 models (4444B-13131313B parameters) across various tasks. (b) Effect of LLM and it’s size on model performance. (c) Average performance of the top-5 models (4444B-13131313B parameters) on multiple-image and single-image questions. | 2501.10360v1 | 6 |
|
Figure B.1:Collage of a subset of images from the dataset, showcasing the diversity of images used in FaceXBench. | 2501.10360v1 | 7 |
|
Figure 2:Overview.Our method jointly predicts pointmaps (X1subscript𝑋1X_{1}italic_X start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT,X2subscript𝑋2X_{2}italic_X start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT) and scene flowS𝑆Sitalic_Swith an information-sharing ViT
backbone followed by three prediction heads (HX1subscriptHsubscript𝑋1\text{H}_{X_{1}}H start_POSTSUBSCRIPT italic_X start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT end_POSTSUBSCRIPT,HX2subscriptHsubscript𝑋2\text{H}_{X_{2}}H start_POSTSUBSCRIPT italic_X start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT end_POSTSUBSCRIPT,HSsubscriptH𝑆\text{H}_{S}H start_POSTSUBSCRIPT italic_S end_POSTSUBSCRIPT). | 2501.10357v2 | 2 |
|
Figure 3:Different datasets have different scales.Here we show a frame from MOVi-F[25], which is in relative scale, and one from Virtual KITTI[11], which is in metric units. We need to account for this while training for both geometry and motion. | 2501.10357v2 | 3 |
|
Figure 5:Qualitative Results: Ours vs. Peers.We compare our method wth theSFpeers and one representative ofΔDΔ𝐷\Delta Droman_Δ italic_D+OFpeers.
We show qualitative results from all datasets.
For each scene we show a pair of input images, and for each peer we show accuracy maps againstOursfor both scene flow (AccR𝐴𝑐𝑐𝑅AccRitalic_A italic_c italic_c italic_R) and depth (δ1subscript𝛿1\delta_{1}italic_δ start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT).
Color legend is shown above. | 2501.10357v2 | 5 |
|
Figure 8:Additional Qualitative Result on DAVIS[71],
with scene flow represented by CIE-LAB[31]colormap overlaid on corresponding 3D structure estimated for first input image.
As compared to other approaches our approach correctly shows the motion on the legs of the dancer, while identifying most of the background as static. | 2501.10357v2 | 8 |
|
Figure 9:Additional Qualitative Result on DAVIS[71],
with scene flow represented by CIE-LAB[31]colormap overlayed on corresponding 3D structure estimated for first input image.
Our scene-flow result correctly identifies the foreground motion as compared to other approaches. | 2501.10357v2 | 9 |
|
Figure 2:Distribution of samples in the training () and test () splits with respect to the ground truth distance | 2501.10343v1 | 6 |
|
Figure 4:Mean IoU over the distance to the objects with a step size of 100 meters for the top 2 submissions compared with the baseline model | 2501.10343v1 | 8 |
|
Figure 5:Distance Estimation Results comparing the top two submissions,Data EnhanceandYolov7 Depth Widened, against the provided baseline model with the absolute error metric (ϵabssubscriptitalic-ϵ𝑎𝑏𝑠\epsilon_{abs}italic_ϵ start_POSTSUBSCRIPT italic_a italic_b italic_s end_POSTSUBSCRIPT). | 2501.10343v1 | 9 |
|
Figure 7:Obstacle detection rate and number of FPs across different sizes of obstacles for top performing methods. | 2501.10343v1 | 15 |
Subsets and Splits