image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 11:Results extending main textFig.5: Efficiency of nonuniform coarse discretizations on Funnel and 25GMM densities. | 2501.06148v1 | 11 |
|
Figure 12:Results extending main textFig.6. Evaluation of models trained withNtrain=10subscript𝑁train10N_{\rm train}=10italic_N start_POSTSUBSCRIPT roman_train end_POSTSUBSCRIPT = 10steps using varying numbers of integration steps. | 2501.06148v1 | 12 |
|
Figure 2:Scaling results on the VoiceBank+Demand dataset. The smallest (N=1𝑁1N=1italic_N = 1) and largest (N=6𝑁6N=6italic_N = 6) models are1.37Mtimes1.37M1.37\text{\mskip 5.0mu plus 5.0mu}\mathrm{M}start_ARG 1.37 end_ARG start_ARG times end_ARG start_ARG roman_M end_ARGand2.94Mtimes2.94M2.94\text{\mskip 5.0mu plus 5.0mu}\mathrm{M}start_ARG 2.94 end_ARG start_ARG times end_ARG start_ARG roman_M end_ARGparameters, respectively. | 2501.06146v1 | 2 |
|
Figure 1:Examples of uploaded problem images: FCI, questions 8-11, in Persian (left panel) and HTCE, questions 16-19, in Chinese (right panel). | 2501.06143v1 | 1 |
|
Figure 7:Heat map and dendrogram of similar language-switching behavior. Dark red indicates strong similarity in language-switching behavior. Blocks around the main diagonal indicate clusters. | 2501.06143v1 | 7 |
|
Figure 2:The activation values for each neuron (denoted by color) at each step
in the trial with a object quantity of 15. Values are averaged over 15 trials.
In the rightmost panel, we label
the specific neurons used in a one-off causal intervention described in
Sections3.5and4.1. | 2501.06141v1 | 2 |
|
Figure 3:(a) and (b) Theoretical neural solutions to the numeric tasks. The purple arrows
represent incoming demo tokens, the darker
arrows indicate the trigger token, the lighter colored arrows indicate increments to the
response tokens, the green dot indicates the starting point. (d) and (e)
show the first two principal components of a
Same-Object and Multi-Object GRUs. Multiple trajectories are shown, each point
is a projected latent state in a trajectory. The lines trace individual trajectories.
(See Appendix17and15for details.)
(c) IIA for the full hidden state substitutions described for the Ctx-Distr program, and the DAS
IIA for the Last Value alignment (see Figure9for
expanded details).
VL stands for the Variable-Length variants of the task in the x-labels.
(f) IIA for the attention interventions. Results from the two layers in each
model seed are sorted based on superior IIA and then averaged over seeds. | 2501.06141v1 | 3 |
|
Figure 4:Interchange intervention accuracy (IIA) on variables from
different symbolic programs for different tasks faceted by
architecture type.
The displayed IIA in the Up-Down program is taken from
the Count variable. The IIA in the Up-Up program
is taken as the better performing of the two possible count variables
for each model type respectively.
All IIA measurements show the proportion of trials in which the model
successfully predicts all counterfactual R and EOS tokens following a
causal intervention. | 2501.06141v1 | 4 |
|
Figure 6:Diagram of the main transformer architecture used in this work. The
white rectangles represent activation vectors. The arrows represent
model operations. Unless
otherwise stated, all interchange interventions were performed on the
Hidden State activations from Layer 1 or the Residual Stream 0 within
Layer 1 for the key and value projections. All normalizations are Layer
Norms(Ba et al.,2016). | 2501.06141v1 | 6 |
|
Figure 7:Left:The model performance on the tasks. This result includes the Multi-Object,
Single-Object, and Same-Object tasks. Each object quantity includes 15
sampled sequences (even when only one configuration exists for that object
quantity). 3 model seeds were dropped from
the LSTM models in the Same-Object task due to lower than 99%
accuracy. One seed was dropped from the transformer models in each
the Single-Object and Same-Object tasks for the same reason.Right:The GRU performance on the tasks facetted by model size (hidden
dimensionality). This result is only for GRUs train on the
Multi-Object task. | 2501.06141v1 | 7 |
|
Figure 8:Interchange intervention accuracy (IIA) on variables from
different symbolic programs for different tasks faceted by
architecture type. The y-axis shows the proportion of trials
in which the model predicts all counterfactual tokens correctly
after a causal intervention for the corresponding variable on
held out data. | 2501.06141v1 | 8 |
|
Figure 10:Attention weights for a single transformer with two layers
using rotary positional encodings trained on the Multi-Object Task.
Queries are displayed on the vertical
axis in order of their appearance starting at the top. Keys are displayed
on the horizontal axis starting from the left. Queries are only able
to attend to themselves and preceding keys. | 2501.06141v1 | 10 |
|
Figure 11:Attention weights for a single transformer with two layers
using rotary positional encodings trained on the Variable-Length variant of the Multi-Object Task. Queries are displayed on the vertical
axis in order of their appearance starting at the top. Keys are displayed
on the horizontal axis starting from the left. Queries are only able
to attend to themselves and preceding keys. | 2501.06141v1 | 11 |
|
Figure 12:Attention weights for a single transformer model seed with two layers and
no positional encodings (NPE) trained on the Multi-Object Task.
Queries are displayed on the vertical
axis in order of their appearance starting at the top. Keys are displayed
on the horizontal axis starting from the left. Queries are only able
to attend to themselves and preceding keys. | 2501.06141v1 | 12 |
|
Figure 13:Left: Attention weights for a single transformer model seed with one layer
and no positional encodings. Right: Attention weights for a single transformer
seed with one layer and no positional encodings trained without the BOS
and trigger
token types. In both figures, queries are displayed on the vertical
axis in order of their appearance in the sequence starting at the top.
Keys are displayed on the horizontal axis starting from the left.
Queries are only able to attend to themselves and preceding keys. | 2501.06141v1 | 13 |
|
Figure 14:Left:The transformer performance on variable length variants of
the 3 tasks.Right:The interchange intervention accuracy using the Ctx-Distr program for the transformer models on the variable length tasks.
In both panels, 4 model seeds were dropped from the models in the
variable length Same-Object task due to lower than 99%
accuracy, and one seed was dropped from the variable length
Single-Object task for the same reason. | 2501.06141v1 | 14 |
|
Figure 15:Principal Components Analysis of a single GRU model seed
including hidden state representations over 10 trials for each
object quantity from 1 to 20 in the Multi-Object task variant.
Green points
indicate the start of a plotted trajectory, black points indicate an
intermediate step, and red points indicate the end of a plotted
trajectory. The blue line plots a single trajectory from start
to finish with a object quantity of 3. Similarly, the orange and
green lines follow single trajectories of 7 and 15 respectively. | 2501.06141v1 | 15 |
|
Figure 16:Principal Components Analysis of a single GRU model seed
including hidden state representations over 10 trials for each
object quantity from 1 to 20 in the Single Object task variant.
Green points
indicate the start of a plotted trajectory, black points indicate an
intermediate step, and red points indicate the end of a plotted
trajectory. The blue line plots a single trajectory from start
to finish with a object quantity of 3. Similarly, the orange and
green lines follow single trajectories of 7 and 15 respectively. | 2501.06141v1 | 16 |
|
Figure 17:Principal Components Analysis of a single GRU model seed
including hidden state representations over 10 trials for each
object quantity from 1 to 20 in the Same-Object task variant.
Green points
indicate the start of a plotted trajectory, black points indicate an
intermediate step, and red points indicate the end of a plotted
trajectory. The blue line plots a single trajectory from start
to finish with a object quantity of 3. Similarly, the orange and
green lines follow single trajectories of 7 and 15 respectively. | 2501.06141v1 | 17 |
|
Figure 2:Overall Architecture ofMultiscale Temporal Mamba (MS-Temba)for action detection. MS-Temba is composed of a Visual Backbone, a Temporal Encoder consisting of Temporal Mamba (Temba) Blocks, a Temporal Mamba Fuser and a Classification Head. | 2501.06138v1 | 2 |
|
Figure 3:Temba Block.A single Temba block consists of a Temporal Local Module (TLM) and a Dilated Temporal SSM. The TLM is responsible for learning local feature while the DTS learns long-range temporal dependencies.ΦΦ\Phiroman_Φdenotes a non-parameterized operation to facilitate dilation in the SSM. | 2501.06138v1 | 3 |
|
Figure 4:Scanning Strategy.An illustration of scanning strategy in Dilated Temporal SSM (DTS). Note this diagram shows only the forward scanning in the second Temba Block. MS-Temba uses bidirectional scanning. | 2501.06138v1 | 4 |
|
Figure 5:Temporal Mamba Fuser (TMF)is composed of projection layers to aggregate features learnt at different Temba Blocks. After these features are fused,BiMambais responsible for interaction between the fused features. | 2501.06138v1 | 5 |
|
Figure 6:Illustration of the evolution of MS-Temba’s accuracy with increase in video duration. MS-Temba’s effectiveness increases considerably with increase in duration. | 2501.06138v1 | 6 |
|
Figure 6:Illustration of the evolution of MS-Temba’s accuracy with increase in video duration. MS-Temba’s effectiveness increases considerably with increase in duration. | 2501.06138v1 | 7 |
|
Figure 7:Throughput vs Video Duration.We show that as video duration increases, the throughput of MS-Temba increases compared to the baseline. | 2501.06138v1 | 8 |
|
Figure 8:Top-5 action improvements.We illustrate the top five actions where our model outperforms the baseline. These include long-duration as well as short actions. | 2501.06138v1 | 9 |
|
(a)Comparison of Information Density of I3D and CLIP features of a sample of 15 frames averaged over 10 videos of Charades | 2501.06138v1 | 15 |
|
(a)Comparison of Information Density of I3D and CLIP features of a sample of 15 frames averaged over 10 videos of Charades | 2501.06138v1 | 16 |
|
(b)Comparison of Information Density of I3D and CLIP features of a sample of 100 frames averaged over 10 videos of TSU | 2501.06138v1 | 17 |
|
Figure 1:(Top) Scenario Projections for Global Evolution of AI Incidents and Hazards (Bottom) Projected Evolution of Incidents and Hazards by AI Principle. | 2501.06137v1 | 1 |
|
Figure 2:Projected Evolution of Jailbreak Prompts with and without the June 2023 Regulatory Intervention. | 2501.06137v1 | 2 |
|
Figure 3:Results of a typical run using the no-prioritisation policy for LLM risk report processing. Proportions are calculated based on the total number of processed reports in the simulation. | 2501.06137v1 | 3 |
|
Figure 4:Results of a typical run using the priority-based policy for LLM risk report processing. Proportions are calculated based on the total number of processed reports in the simulation. | 2501.06137v1 | 4 |
|
Figure 6:The left panel shows the proportion of reports processed from each source (community, crowdsourcing, expert) under different supervision policies. The right panel displays the distribution of processed risk types under each policy. Proportions in both panels are calculated relative to the total number of reports processed for each category. For example, in the left panel, the proportions represent the percentage of processed reports originating from each source under each policy. In the right panel, the proportions show the percentage of reports for each risk type that were processed under a given policy, relative to the total number of reports processed for that risk type. | 2501.06137v1 | 6 |
|
Figure 7:(Top) Incentive Levels Over Time: Expert incentives increase as attention shifts toward expert-driven reports, while community and crowdsourcing incentives decline. (Bottom) Occurrence Rates Over Time: High-priority risks such as security and AI alignment are suppressed, while user experience and content moderation re-emerge. | 2501.06137v1 | 7 |
|
Figure 8:Monthly Simulation Results Overview for the Non-prioritised Policy over the WildChat dataset. Simulation continues beyond the initial 12-month period to allow backlog processing to completion. | 2501.06137v1 | 8 |
|
Figure 9:Monthly Simulation Results Overview for the Prioritised Policy over the WildChat dataset. Simulation continues beyond the initial 12-month period to allow backlog processing to completion. | 2501.06137v1 | 9 |
|
Figure 10:Comparison of the first 10 processed reports each month for bothpriority-basedandnon-prioritisedapproaches. | 2501.06137v1 | 10 |
|
Figure 11:Illustration of priority scorepisubscript𝑝𝑖p_{i}italic_p start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTunder different combinations of accessibilityaisubscript𝑎𝑖a_{i}italic_a start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTand potential damagedisubscript𝑑𝑖d_{i}italic_d start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT, highlighting the logarithmic scaling effect. | 2501.06137v1 | 11 |
|
Figure 12:Illustration of Poisson distributions for report generation rates: community reports (λcom=25subscript𝜆com25\lambda_{\text{com}}=25italic_λ start_POSTSUBSCRIPT com end_POSTSUBSCRIPT = 25), crowdsourced reports (λcrd=12subscript𝜆crd12\lambda_{\text{crd}}=12italic_λ start_POSTSUBSCRIPT crd end_POSTSUBSCRIPT = 12), and expert reports (λexp=5subscript𝜆exp5\lambda_{\text{exp}}=5italic_λ start_POSTSUBSCRIPT exp end_POSTSUBSCRIPT = 5). | 2501.06137v1 | 12 |
|
Figure 14:Priority Distribution of Processed Reports. The plot compares the density of priority scores betweenpriority-basedandnon-prioritisedprocessing. Dashed lines indicate the mean priority for each processing method, highlighting the effectiveness of prioritisation in focusing on critical cases. | 2501.06137v1 | 14 |
|
Figure 1:The dispatching of the CAVs for the passenger requests from the non-fixed taxi stops. All the requests are gradually assigned to the CAVs for personalized transportation services. | 2501.06132v1 | 1 |
|
Figure 2:Overall Architecture of the Proposed CoDriveVLM. This framework encompasses multiple modules, from multimodal input processing to multifunctional output generation of the VLM. The input module consolidates information from traffic infrastructure, passengers, and CAVs for integration by the subsequent multimodal integrator. Within this framework, the VLM performs two primary functions: determining dispatch orders for CAVs in response to passenger requests, and assessing collision risks between CAV pairs to inform subgraph evolution for cooperative motion planning. Additionally, memory and reflection modules facilitate few-shot learning capabilities. | 2501.06132v1 | 2 |
|
Figure 3:Annotation illustration of the BEV image. This image aggregates the information from the road layout, center lines of the lanes, and the states of the traffic participants. The arrows denote the driving direction of the CAVs, and the numbers above each object are their unique ID to facilitate language-based reasoning and inference. | 2501.06132v1 | 3 |
|
Figure 4:Demonstration of the textual dialogue of the VLM agent. The system message is the abstract of the common knowledge of the environment and the task commands for the AMoD services. The human message enhances the understanding of the corresponding BEV image for the VLM agent by supplementing necessary information. The AI message is the textual response of the VLM agent, which is then stored in the memory module as an important resource for the inference of the VLM agent in a new dialog episode. | 2501.06132v1 | 4 |
|
Figure 5:Illustration of the retrieval process for CoDriveVLM. A BEV image and human message for similarity query are embedded with the pre-trained CLIP encoders, before the pair-wise similarity quantification with the memory vectors. The memory items of Top-K𝐾Kitalic_Ksimilarities are retrieved to support in-context learning. | 2501.06132v1 | 5 |
|
Figure 6:The structure of the parallel and cooperative motion planning algorithm. There is no information exchange between any two sets of subgraphsℋ1subscriptℋ1\mathcal{H}_{1}caligraphic_H start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTandℋ2subscriptℋ2\mathcal{H}_{2}caligraphic_H start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT, while the CAVs in each subgraph need to communicate with others iteratively. | 2501.06132v1 | 6 |
|
Figure 7:Demonstration of the environmental settings of the urban scenario and the corresponding service process flow of the AMoD System. | 2501.06132v1 | 7 |
|
Figure 9:The corresponding dialogue for the dispatching decision at the traffic condition at 146.0 s in the simulation trial with 10 CAVs and 30 predefined passenger requests. The AI agent of the CoDriveVLM inferences and reasons with CoT to get the decision of assigning CAV 6 to pick up passenger 17. | 2501.06132v1 | 12 |
|
Figure 10:The corresponding dialogue for the dispatching decision at the traffic condition at 42.0 s in the simulation trial with 15 CAVs and 30 predefined passenger requests. The AI agent of the CoDriveVLM inferences and reasons with CoT to get the decision of assigning CAV 10 to pick up passenger 9. | 2501.06132v1 | 13 |
|
Figure 11:The distance penalty during the cooperative dispatching and motion planning with receding horizon. A smaller distance penalty indicates better safety performance and a smaller collision risk between any pair of CAVs within the time slot. | 2501.06132v1 | 14 |
|
Figure 12:The impact of the memory module and the few-shot number to the performance of the proposed CoDriveVLM framework. The horizontal line is the benchmark value under the zero-shot condition. | 2501.06132v1 | 15 |
|
Figure 1:The overview of our ASR correction pipeline. After intent classification, we re-rankn𝑛nitalic_n-best ASR hypotheses by the lexical and semantic similarity of context retrieved from the dialogue states and the GPT-augmented task list. If none of the hypotheses seems plausible, we find a task or option that sounds phonetically similar to the best hypothesis. | 2501.06129v1 | 1 |
|
Figure 3:ASR & Speech-to-English-Text Translation (S2ETT).CER and sacreBLEU of ASR and S2ETT transcriptions against of Fleurs utterances pool across splits of original Flores sentences. Numbers in parentheses represent the size of the language groups, e.g., Whisper-v3 (SeamlessM4Tv2) supports 84 (89) of 102 Fleurs languages. Neither of the two supports 9 languages. ‘Win Ratio’ displays the number of language a model performs better on the task. Whisper=Whisper-v3-Large;S4T=SeamlessM4Tv2-Large. | 2501.06117v1 | 3 |
|
Figure 3:A screenshot of the preprocessed CMU dataset highlighting its four distinctive features (m, ud, dd, and id). | 2501.07600v1 | 5 |
|
Figure 4:Aalto Dataset: Box plots for both the breadth-wise experiments (as seen horizontally with varying number of subjects), and the depth-wise experiments (as seen vertically with varying number of samples per subject) for 7.6 million triplets, whereM=70𝑀70M=70italic_M = 70. Each box plot displays the EERs from ten reruns. | 2501.07600v1 | 6 |
|
Figure 1:Learning dynamics of the effective model.For an RBM trained with samples of a spin-1 Blume-Capel model with three body interactions defined in Eq. (13) withJ(2)=J(3)=1superscript𝐽2superscript𝐽31J^{(2)}=J^{(3)}=1italic_J start_POSTSUPERSCRIPT ( 2 ) end_POSTSUPERSCRIPT = italic_J start_POSTSUPERSCRIPT ( 3 ) end_POSTSUPERSCRIPT = 1at inverse temperatureβ=0.2𝛽0.2\beta=0.2italic_β = 0.2, (A) shows the evolution of the averaged effective couplings extracted from the RBM along the training timet𝑡titalic_tin both the zero-sum (top) and lattice gas (bottom) gauges. For an RBM trained with the MSA of the PF00072 protein family, (B) displays the Fröbenius norm of the inferred effective model parameters as a function of the training time. We point out the update when we obtain a maximum in the contact prediction performance with the vertical dashed line. In both figures, the solid lines indicate mean values, and the width of the shading areas indicates the standard deviation. | 2501.06108v1 | 1 |
|
Figure 2:Contact Prediction from the MSA.Considering the PF00072 protein family, (A) and (B) show the receiver operating characteristic curve (ROC curve) and the positive predictive value against the number of predictions (PPV curve) at the different stages of the training. (C) display the log-likelihood (LL) concerning the training set and a validation set during the training. In (D), we plotted the area under the ROC and PPV curves versus the training timet𝑡titalic_t. Colored dots and lines in (D) correspond to the different training stages shown in (A) and (B). The models shown here were trained using a PCD-10 scheme, 1000 hidden nodes, and a 0.01 learning rate. | 2501.06108v1 | 2 |
|
Figure 3:Diagram of the 3-visible hidden nodes RBM we use to derive the exact mapping in the zero-sum Gauge. | 2501.06108v1 | 3 |
|
Figure 2.A CPAP setting with3333different states,20202020different types of arms, each type has10101010arms, the budget is set to be10101010and the time Horizon is10101010. | 2501.06103v1 | 2 |
|
Figure 4.We present the average running time ofSPIpolicy, finite whittle policy, and infinite whittle policy in the CPAP setting(N,S,K,ρ,T)=(10,10,50,50,10)𝑁𝑆𝐾𝜌𝑇1010505010(N,S,K,\rho,T)=(10,10,50,50,10)( italic_N , italic_S , italic_K , italic_ρ , italic_T ) = ( 10 , 10 , 50 , 50 , 10 ). | 2501.06103v1 | 4 |
|
Figure 5.We present the chart corresponding to Table2, where the performance of all policies is normalized between the optimal upper bound and the random policy. Specifically, the normalized optimal upper bound is set to 1, while the performance of the random policy is set to 0, providing a clear comparison of the relative performance across all policies. | 2501.06103v1 | 5 |
|
Figure 6.Here we consider the setting(N,S,K,ρ,T)=(20,5,10,10,10)𝑁𝑆𝐾𝜌𝑇205101010(N,S,K,\rho,T)=(20,5,10,10,10)( italic_N , italic_S , italic_K , italic_ρ , italic_T ) = ( 20 , 5 , 10 , 10 , 10 )in the Birth-Death Process(CPAP) domain. We pick20202020different random seeds (00-19191919) and run10101010simulations for MDP generated by each random seed. Here the chart shows the average performance and95%percent9595\%95 %confidence interval of each policy. | 2501.06103v1 | 6 |
|
Figure 7.We present the average running time ofSPIpolicy, finite whittle policy, and infinite whittle policy in three different randomly generated MDPs with(N,S,K,ρ,T)=(10,10,50,50,10)𝑁𝑆𝐾𝜌𝑇1010505010(N,S,K,\rho,T)=(10,10,50,50,10)( italic_N , italic_S , italic_K , italic_ρ , italic_T ) = ( 10 , 10 , 50 , 50 , 10 ). | 2501.06103v1 | 7 |
|
Figure 8.We present the asymptotic optimality of the proposedSPIpolicy with(N,S,K,ρ,T)=(20,10,30,10,6)𝑁𝑆𝐾𝜌𝑇201030106(N,S,K,\rho,T)=(20,10,30,10,6)( italic_N , italic_S , italic_K , italic_ρ , italic_T ) = ( 20 , 10 , 30 , 10 , 6 ). | 2501.06103v1 | 8 |
|
Figure 1:Top: A map of the nine available water quality monitoring sites (red markers) along the River Thames. The geographical visualisations were generated using the Folium Python library[43]with map data sourced from OpenStreetMap[44]. We show the trajectory (a) and PDF (b) of the DO concentration at TCaP over a time span of five years. We apply additive seasonal (c), additive EMD (d), multiplicative seasonal (e) and multiplicative EMD (f) detrending methods. A filtering frequency off=6𝑓6f=6italic_f = 6hours or droppingm=3𝑚3m=3italic_m = 3modes (orange) captures the oscillating trend while preserving the short-term fluctuations. | 2501.07599v1 | 1 |
|
Figure 3:Top: PDFs of oxygen fluctuations obtained via additive seasonal (a), additive EMD (b), multiplicative seasonal (c) and multiplicative EMD (d) detrending methods, for the example of the site TCaP. Regardless of the methods used, detrending leads to non-Gaussian distributions, which can be approximated byq𝑞qitalic_q-Gaussian distributions (purple).
Bottom: For each site along the River Thames, we plot the scale parameterβ𝛽\betaitalic_βand the shape parameterq𝑞qitalic_qfrom theq𝑞qitalic_q-Gaussian fitting against its distance to the Thames Estuary/sea. | 2501.07599v1 | 4 |
|
Figure 4:TPut (top) and TKB (bottom) data plotted against the same-time predictions by LGBM, all in chronological order. The SHAP values are plotted on the right hand side for each site, showing the calculated feature importances. | 2501.07599v1 | 5 |
|
Figure 5:Top: Example windows for the Informer’s forecast oft∈12,48𝑡1248t\in{12,48}italic_t ∈ 12 , 48with input lengths of48484848and192192192192time steps, respectively. The blue line represents the DO input at every time step. While the model processes all features, this visualization only displays the DO. The green dots depict the desired prediction values. The purple crosses represent the forecast of DO made by the Informer model in a single shot. Bottom: A visualization of the attention weights, represented through a heatmap, when predicting 48 future time steps from 192 past steps. Thex𝑥xitalic_x-axis andy𝑦yitalic_y-axis correspond to the keys and queries, respectively, associated with each element in the sequence. Lighter colors correspond to higher weights. | 2501.07599v1 | 6 |
|
Figure 6:By performing a Fast Fourier Transform, we can determine the important frequencies. Note the intuitive peaks at frequencies near 1/year and 1/half day. | 2501.07599v1 | 7 |
|
Figure 7:Illustration of the LGBM model. Decision tree is built leaf by leaf as opposed to other methods such as XGBoost gradient-boosted trees which build level by level. | 2501.07599v1 | 8 |
|
Figure 1:A real-life example of therapist and client exchanges in a problem solving therapy session. | 2501.06101v1 | 1 |
|
Figure 2:Illustration of the sliding window approach for sequence-to-sequence model training with 48 Input sequence and 24 output sequence | 2501.06099v1 | 2 |
|
Figure 5:Global Feature Importance Line Plot of Sequence Windows for LSTM Model on residential dataset | 2501.06099v1 | 5 |
|
Figure 6:SHAP value heatmap for random background dataset selection approach for LSTM Model on residential dataset | 2501.06099v1 | 6 |
|
Figure 7:SHAP value heatmap for similar background dataset selection approach for LSTM Model on residential dataset | 2501.06099v1 | 7 |
|
Figure 8:Density plots of SHAP values for similar vs. random background dataset across 10 features, based on anomalies detected by the LSTM model on residential dataset | 2501.06099v1 | 8 |
|
Figure 1:Model-based decision-making concept presented with the example of an indoor vertical farm. The standard model-based decision-making framework involves estimating a predictive model from data to derive optimal decisions for the real system. Then this predictive model is used in silico to generate the best possible decisions for a given decision-making objective. The model-based decisions are then implemented on the real system. | 2501.06086v1 | 1 |
|
Figure 3:Constructing decision-oriented predictive models: The upper half (Phase 1) of the schematic illustrates the conventional approach to constructing predictive models, focusing on estimating the best fit to the real system’s dynamics using the data and a choice of model estimation method. The lower half (Phase 2) represents the proposed approach to constructing decision-oriented predictive models usingRL. The key difference lies in the estimation approach, where the decision objective is embedded into the estimation process together with the support of the conventional data-fitting approach. | 2501.06086v1 | 3 |
|
Figure 4:(Top row) Illustration of the energy storage exampleV-A. The optimal value function (V⋆superscript𝑉⋆V^{\star}italic_V start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT) and policy (π⋆superscript𝜋⋆\pi^{\star}italic_π start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT) of the MDP are approximated by theMPCasV^⋆superscript^𝑉⋆\hat{V}^{\star}over^ start_ARG italic_V end_ARG start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPTandπ^⋆superscript^𝜋⋆{\hat{\pi}}^{\star}over^ start_ARG italic_π end_ARG start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPT, respectively. Although the value functions are nearly linear, the policies differ significantly due to the activation of the lower bound. (Bottom row) A similar illustration of the non-smooth value function in exampleV-A2. | 2501.06086v1 | 4 |
|
Figure 5:Illustration of the models obtained from the sufficient condition (26) for different values ofΔΔ\Deltaroman_Δ. The left plot shows the resulting value functions for the realMDPand the model-basedMDP(MPC). The right side plot shows the models in terms of𝐬+𝐚𝐬𝐚\boldsymbol{\mathrm{s}}+\boldsymbol{\mathrm{a}}bold_s + bold_a. The solid curve is the expected value of the real state transition, and the other curves are the optimal models for different values ofΔΔ\Deltaroman_Δ. | 2501.06086v1 | 5 |
|
Figure 1:Overview of the unlearnable methods applied to protect private pathology data. The upper section illustrates the traditional pipeline where original private pathology data is used, potentially leading to data breaches. The lower section demonstrates the application of unlearnable methods, which generate unlearnable pathology data to enhance data security and protect against unauthorized model learning. | 2501.06080v1 | 1 |
|
Figure 2:Structure of the DDP-Packed UCs Method, with both generator and surrogate models utilizing DDP.Note that the generator models and surrogate models are distributively deployed on each node of the Summit Supercomputer. | 2501.06080v1 | 2 |
|
Figure 4:Curve plots of model accuracy across batch sizes of 32, 64, 128, 256, and 512 for BloodMNist and PathMNist. Training with 32 batch sizes on PathMNist leads to better performance, while BloodMNist exhibits a similar trend when training with smaller batch sizes(64) but shows weaker training results compared to large batch size variations. This phenomenon supports the notion that the training performance of the UCs under DDP is data-specific and varies depending on the dataset. | 2501.06080v1 | 4 |
|
Figure 5:Bar-plot of the MSE results for before implementing matching (Before), after implementing matching of xFBCI, Ditto, and Individual for two clients. We utilize Client 1 and Client 2 here for representing the two EHD datasets. | 2501.06077v1 | 11 |
|
Figure 1:This figure shows time series of available variables values in the HFID (phase-FEWS, phase-IPC/CH, FCS-LIT, rCSI-LIT, FCS-RT, rCSI-RT) for the sub-region of Ansongo, Gao, Mali. A unique y-axis is used for display, reminding the reader that those variables represent different quantities. | 2501.06076v2 | 1 |
|
Figure 7:The figure illustrates the number of records in agreement between FEWS NET phases (phase-FEWS) in x-axis and IPC/CH phases (phase-IPC/CH) in y-axis. In an ideal scenario of complete agreement between the two classification schemes, the heatmap would form a diagonal matrix. Off-diagonal elements provide insight into areas of disagreement between the two schemes. | 2501.06076v2 | 11 |
|
Figure 8:Box plots representing the distribution population prevalence of insufficient food consumption indicators (FCS, rCSI) from WFP-LIT and from WFP-RT, segmented into five different classes, distinguishing between FEWS NET and IPC/CH classifications. | 2501.06076v2 | 12 |
|
Figure 2:Test inputx𝑥xitalic_x, reference distributionp∗(⋅|x)p^{*}(\cdot|x)italic_p start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT ( ⋅ | italic_x )from the large-scale model, and credal sets produced by CD-CI for small-scale models with different accuracy on the CIFAR-10 data set with classes{airplane, automobile, bird}airplane, automobile, bird\{\text{airplane, automobile, bird}\}{ airplane, automobile, bird }using the KL divergence in (4) with target coverage rate1−ϵ=0.91italic-ϵ0.91-\epsilon=0.91 - italic_ϵ = 0.9. Note that the large-model distributionp∗(⋅|x)p^{*}(\cdot|x)italic_p start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT ( ⋅ | italic_x )is marked as red point in the simplex. | 2501.06066v2 | 4 |
|
Figure 3:Coverage and inefficiency versus the small-scale models accuracy on the CIFAR-10 data set with classes{airplane, automobile, bird}airplane, automobile, bird\{\text{airplane, automobile, bird}\}{ airplane, automobile, bird }using the KL divergence in (4) with target coverage rate1−ϵ=0.91italic-ϵ0.91-\epsilon=0.91 - italic_ϵ = 0.9. The accuracy of the large-scale model, ResNet-18 network, is95.58%percent95.5895.58\%95.58 %, and the accuracy of the small-scale model, Mini-VGG-8, is controlled by training over different numbers of iterations. | 2501.06066v2 | 5 |
|
Figure 8:Confidence histograms for the large-scale model predictive distributionp∗(y|x)superscript𝑝conditional𝑦𝑥p^{*}(y|x)italic_p start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT ( italic_y | italic_x ), for the small-scale modelp(y|x)𝑝conditional𝑦𝑥p(y|x)italic_p ( italic_y | italic_x ), and for CD-CIq∗(y|x)superscript𝑞conditional𝑦𝑥q^{*}(y|x)italic_q start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT ( italic_y | italic_x )evaluated on SNLI data set, usingα=1𝛼1\alpha=1italic_α = 1with target coverage rate1−ϵ=0.91italic-ϵ0.91-\epsilon=0.91 - italic_ϵ = 0.9. | 2501.06066v2 | 10 |
|
Figure 9:Coverage and inefficiency versus the small-scale models accuracy on the SNLI data set using the KL divergence in (4) with target coverage rate1−ϵ=0.91italic-ϵ0.91-\epsilon=0.91 - italic_ϵ = 0.9. The accuracy of the large-scale model is91.58%percent91.5891.58\%91.58 %. Note that INTX𝑋Xitalic_Xdenotes the block-floating point format, in which each element of the block is expressed via INTX𝑋Xitalic_Xprecision (see e.g.,[56,68]). | 2501.06066v2 | 11 |
|
(a)An example showcases MinMo’s capabilities, including speech-to-speech chat, speech-to-text translation, style-controllable speech synthesis, and full duplex interaction. | 2501.06282v1 | 2 |
|
(a)An example showcases MinMo’s capabilities, including speech-to-speech chat, speech-to-text translation, style-controllable speech synthesis, and full duplex interaction. | 2501.06282v1 | 3 |
|
(b)An example showcases MinMo’s capabilities, including speech-to-speech chat, audio event detection, speaker analysis and speech-to-text translation. | 2501.06282v1 | 4 |
|
Figure 3:The overall architecture of MinMo. Table1provides detailed descriptions of each module in this diagram. | 2501.06282v1 | 5 |
|
Figure 4:Detailed training data for the Speech-to-Text Alignment stage.Left:Data distribution forFull-Aligntraining.Right:Data distribution for instruction fine-tuning (SFT). | 2501.06282v1 | 6 |
|
Figure 1:Humans often make sense of new or complex objects by comparing their parts to previously encountered prototypes (Smith et al. (1974)). For example, when describing something unfamiliar, people tend to point out resemblances between parts of the new object and familiar prototypes by stating that ‘this part of the object looks like that other one I have seen before’. We propose a method to classify an image by decomposing it into regions based on learned concepts and tracing each region to the corresponding regions in images from training datasets. We refer to such interpretations as to’COMiX panels’ | 2501.06059v1 | 1 |
|
Figure 3:Examples ofCOMiX panelinterpretations for Oxford-IIIT Pets (left) and CUB-200-211 dataset (right). | 2501.06059v1 | 3 |
Subsets and Splits