image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Comparison between EfficientZero V2, DreamerV3 and other SOTAs in each domain. We evaluate them under the Atari 100k, DMControl Proprio, and DMControl Vision benchmarks.
We then set the performance of the previous SOTA as 1, allowing us to derive normalized mean scores for both EfficientZero V2 and Dreamer V3. EfficientZero V2 surpasses or closely matches the previous SOTA in each domain. | 2403.00564v2 | 1 |
|
Figure 2:Framework of EZ-V2. (A) How EZ-V2 trains its model. The representationℋℋ\mathcal{H}caligraphic_Htakes observations as inputs and outputs the state. The dynamic model𝒢𝒢\mathcal{G}caligraphic_Gpredicts the next state and reward based on the current state and action. Sampling-based Gumbel search outputs the target policyπtsubscript𝜋𝑡\pi_{t}italic_π start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTand target valueztsubscript𝑧𝑡z_{t}italic_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT. (B): How the sampling-based Gumbel search uses the model to plan. The process contains action sampling and selection. The iterative action selection outputs the recommended actionaS∗subscriptsuperscript𝑎𝑆a^{*}_{S}italic_a start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_S end_POSTSUBSCRIPT, search-based value target (target value), and improved policy (target policy). | 2403.00564v2 | 2 |
|
Figure 3:Ablation study of our search method, namely the sampling-based Gumbel search (S-Gumbel search). We compare it with our search method with different numbers of simulations (n=16, 8) and Sample MCTS(Hubert et al.,2021). Our method outperforms Sample MCTS, and increasing the number of simulations improves our method’s performance on hard tasks. | 2403.00564v2 | 3 |
|
Figure 4:Ablation study of our value target, known as the mixed value target. We compare it with different value targets, including the multi-step TD target and the double Q-value target. The mixed value target consistently achieves high performance in both Proprio Control and Vision Control tasks. | 2403.00564v2 | 4 |
|
Figure 5:Intuitive example showing the difference between simple policy loss usingaS∗subscriptsuperscript𝑎𝑆a^{*}_{S}italic_a start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_S end_POSTSUBSCRIPTand cross-entropy loss. | 2403.00564v2 | 5 |
|
Figure 6:DMC scores for proprioceptive inputs with a budget of 200K frames. it corresponds to 100K steps due to the action repeat. (Because different algorithms have varying logging frequencies, the starting points are not the same.) | 2403.00564v2 | 6 |
|
Figure 7:DMC scores for image inputs with a budget of 400K frames. It corresponds to 200K steps due to the action repeat. (Because different algorithms have varying logging frequencies, the starting points are not the same.) | 2403.00564v2 | 7 |
|
Figure 1:CAETraining Instability. For most datasets, theCAEarchitecture exhibits a large spike in reconstruction error that consistently correlates with the unique percentage (definition2.1). | 2403.00563v2 | 1 |
|
Figure 3:IP parametrizations. Validation results for CAE compared against three parametrizations of the linear IP weights on the ISOLET dataset. | 2403.00563v2 | 5 |
|
Figure 4:Varying P. Test set performance on ISOLET for varying size ofIPP𝑃Pitalic_P. The mean reconstruction error with CAE is included as a horizontal line. | 2403.00563v2 | 6 |
|
Figure 2:Overview of the proposed DMTL approach consisting of an early-stage shared feature learning for all the attributes, followed by category-related feature learning for heterogeneous attribute categories. We use the MBConv[11]for shared feature learning. Each task achieves optimal estimation of individual heterogeneous attributes through fine-tuning, e.g., nominal versus ordinal. | 2403.00561v1 | 2 |
|
Figure 7:The proposed approach was deployed and practically applied on Nvidia Jetson Tx2, and software was developed for interaction. | 2403.00561v1 | 22 |
|
Figure 1:Correlations between text diversity scores on CNN/DM. Compression ratio correlates strongly with most other diversity metrics. | 2403.00553v1 | 1 |
|
Figure 2:Correlations between diversity metrics, BERTScore, and ROUGE-1. Both reference-based metrics are weakly correlated with CR and Hom. (BERT), and moderately anti-correlated with Self-BLEU. | 2403.00553v1 | 2 |
|
Figure 3:Mean run time(log-scale)on CNN/DM summaries. Run times increase with the number of text for the analysis. Even for small datasets, Self-BLEU and BERTScore homogenisation are unpractically slow. | 2403.00553v1 | 3 |
|
Figure 1:The reconstruction backbone consists of unrolled gradient descent layers, and the image prior is learned during training by𝒢θsubscript𝒢𝜃\mathcal{G}_{\theta}caligraphic_G start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT. A pre-trained mapping networkℳℳ\mathcal{M}caligraphic_Mis introduced to predict the quantitative parameters𝒑𝒑\bm{p}bold_italic_pand guide the reconstruction with MR relaxometry. | 2403.00549v3 | 1 |
|
Figure 2:Qualitative results for T1 and T2 mapping sequences. The baseline images with the shortest and longest inversion/echo times are shown. The proposed method can generate both images and quantitative maps simultaneously. Perceptually, the reconstructed images of all acceleration factors are of good quality. | 2403.00549v3 | 2 |
|
Figure 1:SURE consistently performs better than previous approaches to uncertainty estimation under various scenarios.Note that we did not manage to scale RegMixup[59]to the learning with noisy label task. Baseline refers to the MSP[31]method. | 2403.00543v1 | 1 |
|
Figure 2:Overview of recipes.Our proposed approach SURE contains two aspects: increasing entropy for hard samples and enforcing flat minima during optimization. We incorporate RegMixup[59]loss and correctness ranking loss (CRL)[54]as our loss function and employ cosine similarity classifier (CSC)[23,33]as our classifier to increase entropy for hard samples. As in optimization, we leverage Sharpness-Aware Minimization (SAM)[19]and Stochastic Weight Averaging (SWA)[35]to find flat minima. | 2403.00543v1 | 2 |
|
Figure 4:The visual results of confidence separation given by different methods on CIFAR100-LT[12]IF=10.SURE leads to better confidence separation than MSP[31]and FMFP[81]. | 2403.00543v1 | 6 |
|
Figure 6:The cantilever beam, its finite-element (FE) mesh, and loading history RH1 for cyclic test[52]. | 2403.00540v3 | 20 |
|
Figure 2:Training of PowerFlowMultiNet: The embedding from Fig.1 is generated with OpenDSS, then concatenated into batches | 2403.00892v3 | 2 |
|
Figure 1:Comparison of poisoned images with their corresponding frequency disparities (amplified by5×5\times5 ×) to clean images of existing attacks.Left: clean images;mid: poisoned images from spatial domain based attacks including BadNets[15], Blend[5], SIG[1], IAB[35]and ReFool[33];right: poisoned images from frequency domain based attacks including FTrojan[49], FIBA[12]and our LFBA attack.
Although state-of-the-art frequency triggers achieve superior perceptual similarity than spatial triggers, they introduce anomaly frequency artifacts. | 2402.15653v2 | 1 |
|
Figure 4:Visualization of network attention by Grad-CAM on GTSRB, CIFAR-10, T-IMNET and CelebA. Compared to the visualization heatmaps of clean images, LFBA does not introduce any unusual regions. | 2402.15653v2 | 13 |
|
Figure 5:Visualization of DCT spectra between clean and posioned samples under various spatial and frequency attacks on two different input-space datasets including CIFAR-10 (32×32323232\times 3232 × 32) and CelebA (64×64646464\times 6464 × 64). We randomly select 10000 samples from each dataset and showcase the averaged spectrum results. | 2402.15653v2 | 14 |
|
Figure 6:Visualization of LFBA poisoned images and triggers under differentϵitalic-ϵ\epsilonitalic_ϵandn𝑛nitalic_n. The pixel value of triggers is amplified by 30×\times×. | 2402.15653v2 | 15 |
|
Figure 1:The Effective Receptive Field (ERF) visualization[46,14]for EDSR[42], RCAN[88], SwinIR[41], HAT[10], and the proposed MambaIR. A larger ERF is indicated by a more extensively distributed dark area. The proposed MambaIR achieves a significant global effective receptive field. | 2402.15648v3 | 1 |
|
Figure 2:The overall network architecture of our MambaIR, as well as the (a) Residual State-Space Block (RSSB), the (b) Vision State-Space Module (VSSM), and the (c) 2D Selective Scan Module (2D-SSM). | 2402.15648v3 | 2 |
|
Figure 3:(a) Without using local enhancement will cause spatially close pixels (area in the red box) get forgotten in the flattened 1D sequence due to the long distance. (b) We use RELU and global average pooling on the VSSM outputs from the last layer to get the channel activation values. Most channels are not activated (i.e., channel redundancy) when channel attention is not used. | 2402.15648v3 | 3 |
|
Figure 4:Qualitative comparison of our MambaIR with CNN and Transformer based methods onclassic image SRwith scale×\times×4. | 2402.15648v3 | 7 |
|
Figure 5:Computational complexity comparison with different input scales. We set the standard attention[65]which has a global receptive field as baseline, and denote it as "Full-attn". We adjust the model to ensure the GPU usage is roughly similar at the beginning, and then scale the input resolution from48×48484848\times 4848 × 48to84×84848484\times 8484 × 84. | 2402.15648v3 | 8 |
|
Figure 6:Qualitative comparison with RealSR[32], ESRGAN[68], BSRGAN[80], Real-ESRGAN[67], and SwinIR[41]onreal image super-resolutionwith scale×\times×4. | 2402.15648v3 | 9 |
|
Figure 7:Qualitative comparison of our MambaIR with other methods oncolor image denoisingtask with noise level levelσ𝜎\sigmaitalic_σ=50. | 2402.15648v3 | 10 |
|
Figure 2:Smoothed weight density plots of classifier layer weights for each predicted class of trojaned defect detection models, poisoned with dead code insertion. | 2402.16896v2 | 2 |
|
Figure 3:Smoothed weight density plots of classifier layer weights for each predicted class of trojaned clone detection models, poisoned with dead code insertion. | 2402.16896v2 | 3 |
|
Figure 4:Smoothed weight density plots of classifier layer weights for each predicted class of trojaned defect detection models that had pretrained weights frozen during finetuning, with variable renaming poisoning. | 2402.16896v2 | 4 |
|
Figure 5:Smoothed weight density plots of classifier layer weights for each predicted class of trojaned models for thedefect detectiontask, poisoned withdead code insertion trigger. | 2402.16896v2 | 5 |
|
Figure 6:Smoothed weight density plots of classifier layer weights for each predicted class of trojaned models for theclone detectiontask, poisoned usingdead code insertiontrigger. | 2402.16896v2 | 6 |
|
Figure 7:Smoothed weight density plots of classifier layer weights for each predicted class of trojaned models for thedefect detection task, poisoned withvariable renaming trigger. | 2402.16896v2 | 7 |
|
Figure 8:Smoothed weight density plots of classifier layer weights for each predicted class of trojaneddefect detectionmodels that hadpretrained weights frozenduring finetuning withdead code insertionpoisoning. | 2402.16896v2 | 8 |
|
Figure 9:Smoothed weight density plots of classifier layer weights for each predicted class of trojaneddefect detectionmodels that hadpretrained weights frozenduring finetuning withvariable renamingpoisoning. (Same as Figure4) | 2402.16896v2 | 9 |
|
Figure 1:The heatmap visualizes the similarities in representations of a specific token within a sample from the last layer outputs across different positions for Llama2-chat-7B. | 2402.15637v2 | 1 |
|
Figure 2:The heat map visualises the similarities in representations of a specific token within a sample from the last encoder layer outputs across different positions for Flan-T5-XL. | 2402.15637v2 | 2 |
|
Figure 3:The overview of our proposedInfoAC. We adopt contrastive learning (Left) to align the representation of a sample,S1subscript𝑆1S_{1}italic_S start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT, as derived from modelM𝑀Mitalic_M, with the representations ofS1subscript𝑆1S_{1}italic_S start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTwhen it is positioned at the end of the sequence derived from a referenced modelMrsubscript𝑀𝑟M_{r}italic_M start_POSTSUBSCRIPT italic_r end_POSTSUBSCRIPT. We also ensure that the hidden representations preceding the classification head are similar when positioned at various locations, resulting in consistent outputs (Right). | 2402.15637v2 | 3 |
|
Figure A1:The heatmap visualizes the similarities in representations of a specific token within a sample from the last layer outputs across different positions for Llama2-chat-7B after fine-tuning withInfoAC. | 2402.15637v2 | 4 |
|
Figure 2:Comparison of ground truth vorticity fields and model predictions for the Navier-Stokes equations example.(a)Time interpolation and extrapolation comparisons in the test set.(b)Error propagation inside (green area) and outside (pink area) the training time domain.(c)The proposed method can conduct arbitrary high-resolution forecasting in space and time. | 2402.15636v1 | 2 |
|
Figure 3:Jerk regularization promotes smoothness and sparsity in latent space.The average jerk is calculated by taking the mean value of the jerk measurements across all snapshots in a trajectory. | 2402.15636v1 | 3 |
|
Figure 4:Comparison of the stage I loss history with and without jerk regularization for a latent space dimension of 10.(a)MSE loss comparison for training and test sets.(b)Jerk loss comparison for training and test sets. | 2402.15636v1 | 4 |
|
Figure 1:PSNR (averaged over 8 images) versus iteration count is depicted for four DIP models fitted to both clean (left panel) and noisy images with noise levelσ=25𝜎25\sigma=25italic_σ = 25(right panel). The 4-layer networks are specified as follows: Blue - kernel size=1, channels [100, 50, 25, 10]; Orange - kernel size=3, channels same as Blue; Green - kernel size=1, channels [128, 128, 128, 128]; Red - kernel size=3, channels same as Green. | 2402.15635v1 | 1 |
|
Figure 2:Newton-Schulz approximation (red) compared with computing exact inverse (purple). Blue, orange and green curves correspond to stopping the update of the inverse after the first5555,10101010, and20202020iterations respectively. | 2402.15635v1 | 2 |
|
Figure 3:(Left) We compare a Bagged-DIP with three sophisticated DIP estimates. (Right) We compare PGD with simple and Bagged-DIPs across different looks on image “Cameraman”. | 2402.15635v1 | 3 |
|
Figure 5:(Left) The structure of Bagged-DIPs withK𝐾Kitalic_Kestimates. (Right) Performance of fitting Bagged-DIP to clean images. | 2402.15635v1 | 5 |
|
Figure 7:Comparison of Bagged-DIPs and three sophisticated DIP estimates on 8 images,m/n=0.5𝑚𝑛0.5m/n=0.5italic_m / italic_n = 0.5,L=50𝐿50L=50italic_L = 50. | 2402.15635v1 | 7 |
|
Figure 8:Comparison between two initializations: (1) proposed: our inititalization method; (2) half: a vector whose elements are all0.50.50.50.5. | 2402.15635v1 | 8 |
|
Figure 1:The example framework of self-endorsement, where only two sampled candidates are leveraged. | 2402.15631v1 | 1 |
|
Figure 5:Step 1 - candidate sampling. We only display 3 candidate samples here and the input prompt is highlighted inblue. | 2402.15631v1 | 15 |
|
Figure 3:Overlapping communication in tensor parallelism (TP) and sequence parallelism (SP) with parallel transformer block (PTB). | 2402.15627v1 | 3 |
|
Figure 6:Inconsistent MFU observed in large-scale training. Different colors denote distinct executions of the same training job. | 2402.15627v1 | 6 |
|
Figure 7:Performance heat-map. The color denotes the running time of the code segments on a rank. The figure also shows the 3D visualization feature, where rank 20 has been selected and the dependency across different parallelism dimensions become visible. | 2402.15627v1 | 7 |
|
Figure 8:The trace shows events collected in a pipeline group on a unified timeline. Dependencies become visible when an event is selected. | 2402.15627v1 | 8 |
|
Figure 9:Weak-scaling training performance of Megatron-LM and MegaScale on the 530B model, where the batch size is scaled proportionally with the number of GPUs. | 2402.15627v1 | 9 |
|
((a))The training loss curve of MegaScale, which includes algorithm optimizations, in comparison with Megatron-LM. | 2402.15627v1 | 10 |
|
((a))The training loss curve of MegaScale, which includes algorithm optimizations, in comparison with Megatron-LM. | 2402.15627v1 | 11 |
|
((b))The training loss curve of ADAM optimizer and LAMB optimizer with four times of the batch size. | 2402.15627v1 | 12 |
|
Figure 11:The normalized training loss curve of a real production run on more than
10,000 GPUs for several weeks. This run trains a model with hundreds of
billions of parameters on multi-trillion tokens. Different colors indicate
training restarts. MegaScale repairs and recovers the training process for
over 100 times in presence of failures. | 2402.15627v1 | 13 |
|
Figure 12:The MFU becomes stable after addressing the stragglers and problematic code segments. Different colors represent different training trials with the same setup. | 2402.15627v1 | 14 |
|
Figure 1:Evolution of the parameters of the estimated adjacency matrix.B∗superscript𝐵∗B^{\ast}italic_B start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPTdenotes the true adjacency matrix andB(n)superscript𝐵𝑛B^{(n)}italic_B start_POSTSUPERSCRIPT ( italic_n ) end_POSTSUPERSCRIPTdenotes the parameters of the estimated adjacency matrix after then𝑛nitalic_n-th training iteration. | 2402.15625v1 | 1 |
|
Figure 2.Fraction of readable predictions for all tasks with different methods and models vs history size. | 2402.15623v1 | 2 |
|
Figure 3.Performance (RMSE, MAE, and error rate) for all tasks with different methods (using Llama 2 13B) vs history size. | 2402.15623v1 | 3 |
|
Figure 4.Bias (mean error) of rating prediction task with different methods and models vs history sizes. | 2402.15623v1 | 4 |
|
Figure 5.Performance metrics (RMSE, MAE, and error rate) for all tasks with different LFM summary lengths with history size 30. | 2402.15623v1 | 5 |
|
Figure 6.Performance (RMSE, MAE, and error rate) for all tasks with LFM (using Llama 2 13B) and NMF using different background sizes with history size 30. | 2402.15623v1 | 6 |
|
Figure 17.RMSE of rating prediction across all approaches and models, with varying user history size. | 2402.15623v1 | 17 |
|
Figure 18.MAE of rating prediction across all approaches and models, with varying user history size. | 2402.15623v1 | 18 |
|
Figure 19.Bias of rating prediction across all approaches and models, with varying user history size. | 2402.15623v1 | 19 |
|
Figure 20.Error of preference prediction across all approaches and models, with varying user history size. | 2402.15623v1 | 20 |
|
Figure 21.Error of choice prediction across all approaches and models, with varying user history size. | 2402.15623v1 | 21 |
|
Figure 1:Active learning with MaxEntropy acquisition function and BERT backbone on QNLI across different strategies over39393939labeling iterations.Left:validation performance after training on labeled data thus far. Error bands represent±1plus-or-minus1\pm 1± 1standard deviation.Right:wall-clock time (in seconds) spent on each phase and validation accuracy of the final model trained on2,00020002,0002 , 000acquired samples. All models trained to convergence on five cores and a Tesla V100-SXM2-32GB GPU. | 2402.15613v1 | 1 |
|
Figure 2:Validation accuracy of final models across different acquisition functions, retraining methods, and datasets. All use BERT as the backbone LLM. Error bands represent±1plus-or-minus1\pm 1± 1standard deviation. | 2402.15613v1 | 2 |
|
Figure 5:Test accuracy of various active learning protocols with MaxEntropy acquisition function and a batch size (bs) of50505050samples per iteration (red-toned curves), or1111sample per iteration (green-toned curves).Top:RoBERTa on SST-2;Bottom:RoBERTa on IMDb. | 2402.15613v1 | 9 |
|
Figure 1:Test 1:Comparison of learned and target trajectories for a random initial condition in[0,1]2dNsuperscript012𝑑𝑁[0,1]^{2dN}[ 0 , 1 ] start_POSTSUPERSCRIPT 2 italic_d italic_N end_POSTSUPERSCRIPT. All the models improve consensus with respect to the uncontrolled dynamics. In terms of learned models, the𝐮VPMPsuperscriptsubscript𝐮𝑉𝑃𝑀𝑃\mathbf{u}_{V}^{PMP}bold_u start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_P italic_M italic_P end_POSTSUPERSCRIPTmodel can be considered the optimal one, with an overall cost (at final time horizonT𝑇Titalic_T)𝒥≈1.44𝒥1.44\mathcal{J}\approx 1.44caligraphic_J ≈ 1.44. | 2402.15611v1 | 1 |
|
Figure 2:Test 1:System configuration at timet=0,1,10𝑡0110t=0,1,10italic_t = 0 , 1 , 10seconds respectively, under the control action provided by the feedback map𝐮VPMPsuperscriptsubscript𝐮𝑉𝑃𝑀𝑃\mathbf{u}_{V}^{PMP}bold_u start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_P italic_M italic_P end_POSTSUPERSCRIPT. | 2402.15611v1 | 2 |
|
Figure 3:Test 2:On the left, comparison of MdPC-controlled and uncontrolled trajectories in terms of particles’ target distance. Variance decay together with updates and variance bounds in the middle (δtol=1subscript𝛿𝑡𝑜𝑙1\delta_{tol}=1italic_δ start_POSTSUBSCRIPT italic_t italic_o italic_l end_POSTSUBSCRIPT = 1) and on the right (δtol=0.1subscript𝛿𝑡𝑜𝑙0.1\delta_{tol}=0.1italic_δ start_POSTSUBSCRIPT italic_t italic_o italic_l end_POSTSUBSCRIPT = 0.1). | 2402.15611v1 | 3 |
|
Figure 4:Test 2:System configuration at timet=0,1,10𝑡0110t=0,1,10italic_t = 0 , 1 , 10seconds respectively, under the control action provided by𝐮MdPCsuperscript𝐮𝑀𝑑𝑃𝐶\mathbf{u}^{MdPC}bold_u start_POSTSUPERSCRIPT italic_M italic_d italic_P italic_C end_POSTSUPERSCRIPTwith toleranceδtol=0.1subscript𝛿𝑡𝑜𝑙0.1\delta_{tol}=0.1italic_δ start_POSTSUBSCRIPT italic_t italic_o italic_l end_POSTSUBSCRIPT = 0.1. | 2402.15611v1 | 4 |
|
Figure 1:Illustration ofReCoVERR.
The VLM predicts that the floor has two tile colors with low confidence.
Instead of abstaining,ReCoVERRcollects reliable and relevant visual evidences related to the question.ReCoVERRmakes salient the evidence that the floor tiles are red and white, helping to verify the VLM’s original answer. | 2402.15610v2 | 1 |
|
Figure 2:TheReCoVERRalgorithm. If the VLM is uncertain in its prediction (1),ReCoVERRtries to verify the VLM hypothesis by collecting evidences.ReCoVERRundertakes multiple turns of evidence collection, which involves generating visual evidences by using an LLM to ask questions to the VLM (2), retaining the reliable (3) and relevant (4) evidences, and checking whether the collected evidence entails the hypothesis (5). | 2402.15610v2 | 2 |
|
Figure 1:The whole network architectures of Algorithm1: (a) PI-FC; (b) PI-LSTM. The data in the green boxes will used to calculate the loss. The solid lines represent the data flow generated in the current iteration. AutoD is short of the automatic differentiation. SDE means the Euler-Maruyama scheme. | 2402.15592v1 | 1 |
|
Figure 2:The whole network architectures of Algorithm2: (a) PI-FC; (b) PI-LSTM. The data in the green boxes will used to calculate the loss. The solid lines represent the data flow generated in the current iteration. AutoD is short of the automatic differentiation. SDE means the Euler-Maruyama scheme. | 2402.15592v1 | 4 |
|
Figure 3:Numerical results of Example 1. Left: state trajectory; Middle: neural optimal controller; Right: the value function. | 2402.15592v1 | 7 |
|
Figure 4:Numerical results of low-dimensional linear control problems. (a)n=2𝑛2n=2italic_n = 2. Left: position𝐱𝐱\mathbf{x}bold_x; Right: control𝐮𝐮\mathbf{u}bold_u. (b)n=4𝑛4n=4italic_n = 4. Upper: based on FC; Below: based on LSTM. | 2402.15592v1 | 8 |
|
Figure 5:Numerical results of30303030-dimensional linear control problems based on LSTM. Left: position𝐱𝐱\mathbf{x}bold_x; Right: control𝐮𝐮\mathbf{u}bold_u. | 2402.15592v1 | 11 |
|
Figure 6:Numerical results of controlled pendulum system. Upper left: pendulum angle; Upper right: pendulum rate; Below left: the control; Below right: the value function. | 2402.15592v1 | 12 |
|
Figure 7:Numerical results of controlled cart-ploe system based on LSTM. Upper left: cart position; Upper middle: cart velocity; Upper right: pole angle; Below left: pole rate; Below middle: the control; Below right: the value function. | 2402.15592v1 | 13 |
|
Figure 8:Numerical results of controlled planner quadcopter system based on LSTM. (a) Upper left: positionx𝑥xitalic_x; Upper middle: positiony𝑦yitalic_y; Upper right: angle; Below left: velocity alongx𝑥xitalic_x-axis; Below middle: velocity alongy𝑦yitalic_y-axis; Below right: angular velocity. (b) Left: the trajectory; Middle: the control; Right: the value function. | 2402.15592v1 | 14 |
|
Figure 9:The values of the loss with varying parameter of the time interval partition. Left: the loss of Algorithm1implemented by LSTM in Example 1. Right: the loss of Algorithm2implemented by LSTM in Example 2 withn=2𝑛2n=2italic_n = 2. | 2402.15592v1 | 17 |
|
Figure 1:(a)The RecWizard architecture comprises pipeline and module levels. Text data flows between modules and the pipeline, while tensor data flows within modules after being processed by RecWizard tokenizers, ensuring RecWizard modularity and portability.(b)In INFO mode, the RecWizard exampleChatGPT-expansionincludes the ReDIAL-Rec recommender module and ChatGPT with prompts as the generator module. Users can select models, set basic arguments, and chat with RecWizard.(c)In DEBUG mode, using theUniCRS-ReDIALmodel as an example, module-level timeline visualization and intermediate messages are enabled for debugging or in-depth demonstrations in the user interface. | 2402.15591v1 | 1 |
|
Fig. 1:TELeR Taxonomy for prompting LLMs to perform complex tasks. For details, seeSantu and Feng (2023). | 2402.15589v1 | 1 |
|
Figure 1:Are LLMs equally aware of social artifacts from different subcultures within a country? For social artifacts from different state-based subcultures of India, we prompt LLMs with unique information that differentiates a social artifact from others and evaluate their overall accuracy for each state. | 2403.14651v1 | 1 |
|
(a)Illustration of the game mechanics and the knowledge elicited from the participants (from Punjab). | 2403.14651v1 | 2 |
|
(a)Illustration of the game mechanics and the knowledge elicited from the participants (from Punjab). | 2403.14651v1 | 3 |
Subsets and Splits