tok2000/arxivcap2k · Datasets at Hugging Face

text stringlengths 9 1.99k	image imagewidth (px) 384 384
We train agents to perform in visually complex Gibson <cit.> and <cit.> environments such as the ones shown here. These environments feature detailed scans of real-world scenes composed of up to 600K triangles and high-resolution textures. Our system is able to train agents using 64$\times$64 depth sensors (a high-resolution example is shown on the left) in these environments at 19,900 frames per second, and agents with 64$\times$64 RGB cameras at 13,300 frames per second on a single GPU.
The batch simulation and rendering architecture. Each component communicates at the granularity of batches of $N$ elements (e.g., $N$=1024), minimizing communication overheads and allowing components to independently parallelize their execution over each batch. To fit the working set for large batches on the GPU, the renderer maintains $K\! \ll\! N$ unique scene assets in GPU memory and shares these assets across subsets of the $N$ environments in a batch. To enable experience collection across a diverse set of environments, the renderer continuously updates the set of $K$ in-memory scene assets using asynchronous transfers that overlap rollout generation and learning.
SPL vs. wall-clock time (agents) on a RTX 3090 over 48 hours (time required to reach 2.5 billion samples with ). exceeds $80\%$ SPL in 10 hours and achieves a significantly higher SPL than the baselines. SPL vs. wall-clock time (training agents over 2.5 billion samples on 8 Tesla V100s) for various batch sizes ($N$). $N$=256 finishes after 2$\times$ the wall-clock time as $N$=1024, but both achieve statistically similar SPL.
SPL vs. wall-clock time (agents) on a RTX 3090 over 48 hours (time required to reach 2.5 billion samples with ). exceeds $80\%$ SPL in 10 hours and achieves a significantly higher SPL than the baselines. SPL vs. wall-clock time (training agents over 2.5 billion samples on 8 Tesla V100s) for various batch sizes ($N$). $N$=256 finishes after 2$\times$ the wall-clock time as $N$=1024, but both achieve statistically similar SPL.
's validation set SPL for vs. number of training samples across a range of batch sizes. This graph shows that sample efficiency slightly decreases with larger batch sizes (with the exception of $N$=512 vs. $N$=1024, where $N$=1024 exhibits better validation score). Ultimately, the difference in converged performance is less than 1% SPL between different batch sizes. Although $N$=256 converges the fastest in terms of training samples needed, {fig:batchspl} shows that $N$=256 performs poorly in terms of SPL achieved per unit of training time. Frames per second achieved by the standalone renderer on a RTX 3090 across a range of resolutions and batch sizes for a sensor on the Gibson dataset. Performance saturates at a batch size of 512. For lower batch sizes, increasing resolution has a minimal performance impact, because the GPU still isn't fully utilized. As resolution increases with larger batches, the relative decrease in performance from higher resolution increases.
's validation set SPL for vs. number of training samples across a range of batch sizes. This graph shows that sample efficiency slightly decreases with larger batch sizes (with the exception of $N$=512 vs. $N$=1024, where $N$=1024 exhibits better validation score). Ultimately, the difference in converged performance is less than 1% SPL between different batch sizes. Although $N$=256 converges the fastest in terms of training samples needed, {fig:batchspl} shows that $N$=256 performs poorly in terms of SPL achieved per unit of training time. Frames per second achieved by the standalone renderer on a RTX 3090 across a range of resolutions and batch sizes for a sensor on the Gibson dataset. Performance saturates at a batch size of 512. For lower batch sizes, increasing resolution has a minimal performance impact, because the GPU still isn't fully utilized. As resolution increases with larger batches, the relative decrease in performance from higher resolution increases.
The effect of the Lamb optimizer versus the baseline Adam optimizer on sample efficiency while training a sensor driven agent. Lamb maintains a consistent lead in terms of SPL throughout training, especially in the first half of training.
We run five trials of AP between $\mathcal A$ and $\mathcal B^{\text{lb}}$ with random initializations, where $N = 10$ and $R = 10$. For each trial, we plot the ratios $d(a_{k+1},E)/d(a_k,E)$, where $E =\mathcal A \cap \mathcal B^{\text{lb}}$ is the optimal set. The red line shows the theoretical lower bound of $1 - \tfrac{1}{R}(1-\cos(\tfrac{2\pi}{N}))$ on the worst-case rate of convergence.
Relative frequency errors $\tilde{e}_\infty (G, G_d^k)$, $\tilde{e}_\infty(G, P_\infty(G_d^k))$ versus the order $k$. A star marker indicates an unstable model.
Frequency-domain responses (up to the Nyquist frequency) of the continuous-time model $G$ and its discretised counterparts $G_d^{zoh}$, $G_d^{ii}$ and $G_d$.
Impulse responses of $G$ and its discretised counterparts $G_d^{zoh}$ (top), $G_d^{ii}$ (middle) and $G_d$ (bottom).
Time-domain error $y(t) - y_s(t)$ associated with the impulse responses of $G_d^{zoh}$, $G_d^{ii}$ and $G_d$.
Relative frequency errors $\tilde{e}_\infty (G, G_d^k)$, $\tilde{e}_\infty(G, P_\infty(G_d^k))$ versus the order $k$ for the TDS case (<ref>).
Frequency-domain responses of the time-delay system (<ref>) and its discretised counterparts.
Step responses of the time-delay system (<ref>) and its discretised counterparts $G_d^{tus}$ (top) and $G_d$ (bottom).
MiniCrafter environment: Example state from the environment. The water represents the traps, the $3$ resource types wood, iron, coal are present and a single crafting table.
Comparison of different SF pre-training experiments. HTR means Hindsight Task Replacement, TR means Task Replacement, $n$ and $1$ mean whether we used as many policies as features or a single one. The plots show the cumulative task competition on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Comparison of different SF pre-training experiments. HTR means Hindsight Task Replacement, TR means Task Replacement, $n$ and $1$ mean whether we used as many policies as features or a single one. The plots show the cumulative task competition on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Comparison of different SF pre-training experiments. HTR means Hindsight Task Replacement, TR means Task Replacement, $n$ and $1$ mean whether we used as many policies as features or a single one. The plots show the cumulative task competition on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Comparison of different SF pre-training experiments. HTR means Hindsight Task Replacement, TR means Task Replacement, $n$ and $1$ mean whether we used as many policies as features or a single one. The plots show the cumulative task competition on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Pre-training experiments. Comparison of the best SF (Task Replacement with $n$ policies) against baselines. The plots show the cumulative task completion on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Pre-training experiments. Comparison of the best SF (Task Replacement with $n$ policies) against baselines. The plots show the cumulative task completion on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Pre-training experiments. Comparison of the best SF (Task Replacement with $n$ policies) against baselines. The plots show the cumulative task completion on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Pre-training experiments. Comparison of the best SF (Task Replacement with $n$ policies) against baselines. The plots show the cumulative task completion on each task during training: Collect wood, Collect iron, Collect string and Collect table.
Left side shows the performance of the best SF (standard SF with n policies in this case) compared to the baselines on the task "one_item". Middle shows the comparison of all SF training methods. Right side shows the comparison of all SF and baselines on the task "two_item".
Left side shows the performance of the best SF (standard SF with n policies in this case) compared to the baselines on the task "one_item". Middle shows the comparison of all SF training methods. Right side shows the comparison of all SF and baselines on the task "two_item".
Left side shows the performance of the best SF (standard SF with n policies in this case) compared to the baselines on the task "one_item". Middle shows the comparison of all SF training methods. Right side shows the comparison of all SF and baselines on the task "two_item".
Experiments on the "random" and "random_pen" targets. The plots show the running mean reward and the standard error (shaded area) during training. From left to right, the first and the third plots show the comparison of the best SF agent against the baseline agents, while the second and the final plots show the comparison of the SF agents with the different training methods used in this paper.
Experiments on the "random" and "random_pen" targets. The plots show the running mean reward and the standard error (shaded area) during training. From left to right, the first and the third plots show the comparison of the best SF agent against the baseline agents, while the second and the final plots show the comparison of the SF agents with the different training methods used in this paper.
Experiments on the "random" and "random_pen" targets. The plots show the running mean reward and the standard error (shaded area) during training. From left to right, the first and the third plots show the comparison of the best SF agent against the baseline agents, while the second and the final plots show the comparison of the SF agents with the different training methods used in this paper.
Experiments on the "random" and "random_pen" targets. The plots show the running mean reward and the standard error (shaded area) during training. From left to right, the first and the third plots show the comparison of the best SF agent against the baseline agents, while the second and the final plots show the comparison of the SF agents with the different training methods used in this paper.
Left side shows the comparison of all SF training settings on the task "craft_staff". The other plots show all SFs compared to the baselines on all the crafting tasks.
Left side shows the comparison of all SF training settings on the task "craft_staff". The other plots show all SFs compared to the baselines on all the crafting tasks.
Left side shows the comparison of all SF training settings on the task "craft_staff". The other plots show all SFs compared to the baselines on all the crafting tasks.
Left side shows the comparison of all SF training settings on the task "craft_staff". The other plots show all SFs compared to the baselines on all the crafting tasks.
Adversarial body shape search
Averages of the average cumulative rewards with attack strength $\epsilon=0.0,0.0.1,0.05,0.1$ for Walker2d-2d
Averages of the average cumulative rewards with attack strength $\epsilon=0.0,0.0.1,0.05,0.1$ for Ant-2d
Averages of the average cumulative rewards with attack strength $\epsilon=0.0,0.0.1,0.05,0.1$ for Humanoid-2d
Adversarial body shapes with length and thickness perturbations for Walker2d-v2.
Adversarial body shapes with length and thickness perturbations for Walker2d-v2.
Adversarial body shapes with length and thickness perturbations for Walker2d-v2.
Adversarial body shapes with length and thickness perturbations for Ant-v2.
Adversarial body shapes with length and thickness perturbations for Ant-v2.
Adversarial body shapes with length and thickness perturbations for Ant-v2.
Adversarial body shapes with length and thickness perturbations for Humanoid-v2.
Adversarial body shapes with length and thickness perturbations for Humanoid-v2.
Adversarial body shapes with length and thickness perturbations for Humanoid-v2.
{Coupling strength between dynamical quarks and the gluonic sector quantified through the center symmetry breaking $\Delta$ as a function of dimensionless curvature $R/T^2$.}
Illustration of the process of mapping the Lorentzian generators to the Euclidean generators. We start with quantizing the theory with respect to constant $P^0-K^0$ slices on the Lorentzian plane. Those are mapped to constant time slices on the Lorentzian cylinder. The Wick rotation maps those to constant time slices on the Euclidean cylinder. Finally we map the Euclidean cylinder to the Euclidean plane via a radial map.
Illustration of two nearby timelike geodesics in $\mathrm{AdS}_3$ (blue, red) corresponding to two boundary circuits and the minimal (green) and maximal (brown) perpendicular distance between them. The infinitesimal variation was exaggerated to improve the visualization.
Illustration of two nearby timelike geodesics in $\mathrm{AdS}_3$ (blue, red) corresponding to two boundary circuits and the minimal (green) and maximal (brown) perpendicular distance between them. The infinitesimal variation was exaggerated to improve the visualization.
Illustration of two nearby timelike geodesics in $\mathrm{AdS}_3$ (blue, red) corresponding to two boundary circuits and the minimal (green) and maximal (brown) perpendicular distance between them. The infinitesimal variation was exaggerated to improve the visualization.
Illustration of the test statistic and the calculation of the $p$-value of the null hypothesis.
$P$-value (color bars; in unit of %) of the null hypothesis (i.e. the multi-band light curves are consistent with Doppler boost plus DRW model) as a function of the DRW parameters $\sigma$ and $\tau$, with $A_{\rm DB}$ = 1 (top left), 1.5 (top right), 2.17 (middle left), 3 (middle right), 4 (bottom left) and 5 (bottom right) and UV/optical noise ratio $\sigma_{\rm UV}/\sigma_{\rm opt}=1$, considering the {Swift} $V$- and $M2$-bands. The black solid line represents the 5% $p$-value threshold, which separates models that are rejected (red) from those passing the test (blue). The star corresponds to the average $\sigma$ and $\tau$ for quasars with properties similar to PG1302-102 from <cit.>.
$P$-value (color bars; in unit of %) as a function $\sigma$ and $\tau$, for $A_{\rm DB}=2.17$ and with $r_{\rm noise}=\sigma_{\rm UV}/\sigma_{\rm opt}=1$ (top left), 2 (top right), 3 (bottom left) and 4 (bottom right), again considering the {Swift} $V$- and $M2$-bands.
UV spectra from HST (black) and GALEX (grey and light grey). The blue lines show power-law fits to the continuum. The transmission curves of the {Swift}/UVOT (and GALEX NUV) filters are shown to delineate the wavelength coverage of each band - the transmission curves are for illustrative purposes and are not measured in flux units ($y$-axis). The shaded band indicates the wavelength range in which the near-UV spectral slope was estimated in <cit.>.
Optical spectra taken with DBSP at Palomar and LRIS at Keck.
$P$-value (color bars; in unit of %) as a function of DRW parameters $\sigma$ and $\tau$ for six combinations of independent UV ($M2$ and $W1$) and optical ($V$ and $B$) bands, where $A_{\rm DB}$ = 2.17 in panels (a)-(d) and $A_{\rm DB}$ = 1 in (e) and (f), and UV/optical noise ratio $r_{\rm noise}=1$.
Comparison between ASAS-SN and {Swift} $V$-band photometry. The ASAS-SN data points are identical in both panels and are shown in light grey, except for the points taken closest to the time of the nine {Swift} observations analysed in this paper, which are shown in black. The red data points show {Swift} photometry from co-added images (top panel) and from the longest-exposure single image (bottom panel). The co-added {Swift} data are in better agreement with ASAS-SN.
Optical $V$-band (top panel) and near-UV $M2$-band (bottom panel) light curve of PG1301-102, similar to Figure <ref>. The prediction of the Doppler model with parameters from is shown with a solid light-blue line. The dashed dark-blue line represents a sinusoidal fit of the extended optical light curve, with period P=2095 d longer than in , and the corresponding Doppler boost prediction for the UV data with $A_{\rm DB}=2.17$ in the bottom panel.
$P$-values inferred when the fiducial model is fit to three different hypothetical future datasets with extended baselines. The red triangle denotes the original 9 {Swift} points. The dark blue curve is inferred from mock data generated in the fiducial model itself. The light blue curve assumes that the true light curve has a larger Doppler amplitude ratio ($A_{DB}=3$). The dark red curve corresponds to future mock data consisting of pure DRW variability.
Top Panel: Optical light curve of PG1302-102 with data from <cit.> in black (CRTS+LINEAR and other archival observations), data from ASAS-SN in grey, and purple squares/red diamonds for {Swift} $B$/$V$-band observations. Bottom panel: Near-UV light curve, with black circles and triangles for {GALEX} and {HST} observations from , purple squares and red diamonds for {Swift} $W1$ and $M2$-band observations, respectively. The sinusoidal Doppler boost model from is also shown in light blue.
$P$-values (color bars; in unit of %) of three pairs of independent combinations of bands {vs} the DRW parameters, assuming Doppler amplitude $A_{\rm DB}$ = 1 for the combinations of bands in panel (a) and $A_{\rm DB}$ = 2.17 in panels (b) and (C), and UV/optical noise ratio $r_{\rm noise}=1$.
Example hypothetical light curves over upcoming $\sim$10 years, for each of the three cases explained in <ref>; (1) the fiducial DB and DRW model with $A_{\rm DB}$=2.17 dominates the variability of PG1302 (left); (2) DB + DRW model has higher Doppler amplitude ratio, $A_{\rm DB}$=3 (middle); and (3) DB is absent from the system and DRW model dominates the true variability (right). In all three scenarios, the DRW amplitude ratio is fixed at 1 (i.e. $r_{\rm noise}$=1). The black points with error bars are the existing 9 {Swift} observations in optical (top panels) and UV (bottom panels). The red and dark blue points with error bars are the hypothetical optical and UV data, respectively.
Example hypothetical light curves over upcoming $\sim$10 years, for each of the three cases explained in <ref>; (1) the fiducial DB and DRW model with $A_{\rm DB}$=2.17 dominates the variability of PG1302 (left); (2) DB + DRW model has higher Doppler amplitude ratio, $A_{\rm DB}$=3 (middle); and (3) DB is absent from the system and DRW model dominates the true variability (right). In all three scenarios, the DRW amplitude ratio is fixed at 1 (i.e. $r_{\rm noise}$=1). The black points with error bars are the existing 9 {Swift} observations in optical (top panels) and UV (bottom panels). The red and dark blue points with error bars are the hypothetical optical and UV data, respectively.
Example hypothetical light curves over upcoming $\sim$10 years, for each of the three cases explained in <ref>; (1) the fiducial DB and DRW model with $A_{\rm DB}$=2.17 dominates the variability of PG1302 (left); (2) DB + DRW model has higher Doppler amplitude ratio, $A_{\rm DB}$=3 (middle); and (3) DB is absent from the system and DRW model dominates the true variability (right). In all three scenarios, the DRW amplitude ratio is fixed at 1 (i.e. $r_{\rm noise}$=1). The black points with error bars are the existing 9 {Swift} observations in optical (top panels) and UV (bottom panels). The red and dark blue points with error bars are the hypothetical optical and UV data, respectively.
Unimodal representation of a single modality can be either effective or not. The effectiveness of different unimodal representations from the same sample also varies. To empower the interaction between modalities, our proposed method aligns the unimodal representation to the effective modality sample-wise and makes full use of the effective unimodal representation under the supervision of the unimodal prediction (F and T represent correct and incorrect predictions, respectively).
The relationship comparison between two modalities in training mini-batch of (a) unsupervised MMC, (b) supervised MMC and (c) UniS-MMC.
Unimodal representation distribution of the first 10 categories of the N24News test set across different methods: (a) aggregation-based method, (b) unsupervised multimodal method, (c) supervised contrastive method and (d) unimodality-supervised method.
Multimodal representation distribution of the first 10 categories of the N24News test set across different methods: (a) aggregation-based method, (b) unsupervised multimodal method, (c) supervised contrastive method and (d) unimodality-supervised method.
As the training progresses, the change of the proportion of both wrong (left), both correct (right) unimodal predictions of the validation set (N24News): the complete method (UniS-MMC), remove negative pair (w.o. C_{Neg}), remove semi-positive pair (w.o. C_{Semi}) and remove both (w.o. C_{Neg},C_{Semi}).
Consistency comparison of unimodal prediction between MT-MML and the UniS-MMC.
The framework for our proposed UniS-MMC.
Loss function for L1 estimator Loss function for L2 estimator
Loss function for L1 estimator Loss function for L2 estimator
Estimation for three-parameter model.
Estimation for one-parameter model.
(a) A DL model's screening accuracy in different label uncertainty groups; (b) GON severity distributions on different label uncertainty. Both x-axes are label uncertainty scores.
Image examples with different GON severity levels and empirical uncertainty scores from the GON dataset.
Overview of the training phase and the inference phase of the uncertainty-guided multi-stream screening model.
Computed margins $\lambda^j_t$, effective disturbances $\hat{\delta}^j_t$ and disturbance $w^j_t$ for every node $x^j$. $\mu_t = \sum\limits_{k} \left\\| \Delta_t(k)\right\\|_1$ is computed with the real system and the controller at $t$.
State and input trajectories for closed loop simulation of controller {algo:DLAR} with perfect parameter knowledge $\alpha$.
Left: Overlay of projections of $\Pt{t}^j$ onto different coordinates for different time steps. Each row corresponds to a different subsystem $x_1$, $x_3$, $x_5$ (top to bottom). Shading indicates time of computation with shades lightening as simulation time passes. Right: state and input trajectories of closed loop simulation with {algo:DLAR} and uncertainties on $\alpha$ described in {sec:introex}.
Left: Overlay of projections of $\Pt{t}^j$ onto different coordinates for different time steps. Each row corresponds to a different subsystem $x_1$, $x_3$, $x_5$ (top to bottom). Shading indicates time of computation with shades lightening as simulation time passes. Right: state and input trajectories of closed loop simulation with {algo:DLAR} and uncertainties on $\alpha$ described in {sec:introex}.
(A) The initial bipartite network structure. Each ego is connected to $2$ alters out of $6$. (B) Protocol for each of the $5$ rounds. In turn-1, an ego (blue) generates ideas independently. In turn-2, they view the ideas of the two alters (red) they are following, and submit any new ideas inspired by the stimuli. Then, they rate the ideas of all $6$ alters, and update which two alters to follow in the next round. (C) The $4$ study conditions. The top, middle and bottom tiers of alters, as recorded from C1, are respectively shown in teal, mustard, and orange bars. (D) Non-redundant idea counts across conditions. Whiskers denote $95\%$ C.I.
Examples of different TST. Including sentiment style transfer (negative $\leftrightarrow$ positive), formality style transfer (informal $\leftrightarrow$ formal), code-switch style transfer (single language $\leftrightarrow$ code-switch sentence).
Comparison of style transfer outputs of our models and style transformer in three transfer tasks. Our models are stage II models in Table <ref>. Translations for code-switching sentences are shown in parenthesis.
Illustration of style weight $w$ vs. Acc, PPL and BLEU in sentiment, formality and code-switching transfer tasks. Note there is a trade-off between Acc and PPL/BLEU. With increasing of $w$, Acc will increase while BLEU drops down and PPL increases. sentiment transfer
Illustration of style weight $w$ vs. Acc, PPL and BLEU in sentiment, formality and code-switching transfer tasks. Note there is a trade-off between Acc and PPL/BLEU. With increasing of $w$, Acc will increase while BLEU drops down and PPL increases. formality transfer
Illustration of style weight $w$ vs. Acc, PPL and BLEU in sentiment, formality and code-switching transfer tasks. Note there is a trade-off between Acc and PPL/BLEU. With increasing of $w$, Acc will increase while BLEU drops down and PPL increases. code-switching transfer
Illustration of the mean attention weights of token `$<$s$>$' from all heads at the final layer in three TST tasks. Higher importance scores are assigned to pivot words, which are depicted as deeper lines in the figures. sentiment transfer
Illustration of the mean attention weights of token `$<$s$>$' from all heads at the final layer in three TST tasks. Higher importance scores are assigned to pivot words, which are depicted as deeper lines in the figures. formality transfer
Illustration of the mean attention weights of token `$<$s$>$' from all heads at the final layer in three TST tasks. Higher importance scores are assigned to pivot words, which are depicted as deeper lines in the figures. code-switching transfer
A single-cell MRC-WPT network with $K$ receivers.
Outage probability in the strong coupling region; Markers and dashed lines correspond to simulations and analytical results, respectively.
Outage probability in the loosely coupling region; Markers and dashed lines correspond to simulations and analytical results, respectively.
Harvested power at each receiver with different loads.
Atomic levels and phase matching in $\Lambda$-scheme Raman scattering induced by classical laser field $\mathcal{E}_{\mathrm{W}}$ and $\mathcal{E}_{\mathrm{R}}$. (a) In the spontaneous write-in process Stokes photons ($\protect\ad_{\mathrm{WS}}$ - mode) and spin-wave excitation ($\protect\bd$ - mode) are created pairwise. (b) Four-wave mixing in readout consists in simultaneous anti-Stokes and Stokes scattering into modes $\protect\ad_{\mathrm{RA}}$ and $\protect\ad_{\mathrm{RS}}$. (c) Phase matching condition or momentum conservation dictates the wave vectors of single photons coupled to a spin-wave excitation with a certain wave vector $\mathbf{K}_{b}$. Write beam with $\mathbf{k}_{\mathrm{W}}$ wave vector is scattered as Stokes photon with $\mathbf{k}_{\mathrm{WS}}$ wave vector, while read beam with $\mathbf{k}_{\mathrm{R}}$ wave vector either scatters Stokes or couples anti-Stokes of the respective wave vectors $\mathbf{k}_{\mathrm{RS}}$, $\mathbf{k}_{\mathrm{RA}}$.
(a) to (c) - temporal evolution of gains $G_{i}$ and spontaneous noises $S_{i}$ building up the light fields. Data plotted for different values of coupling coefficients $\chi$ and $\xi$ correspond to detuning parameters $\Delta_{R}$= 0.3, 0.6 and 0.8. (d) Integrated readout gains $\bar{G}_{i}$ and spontaneous noises $\bar{S}_{i}$ versus detuning $\Delta_{R}$.

We train agents to perform in visually complex Gibson <cit.> and <cit.> environments such as the ones shown here. These environments feature detailed scans of real-world scenes composed of up to 600K triangles and high-resolution textures. Our system is able to train agents using 64$\times$64 depth sensors (a high-resolution example is shown on the left) in these environments at 19,900 frames per second, and agents with 64$\times$64 RGB cameras at 13,300 frames per second on a single GPU.

The batch simulation and rendering architecture. Each component communicates at the granularity of batches of $N$ elements (e.g., $N$=1024), minimizing communication overheads and allowing components to independently parallelize their execution over each batch. To fit the working set for large batches on the GPU, the renderer maintains $K\! \ll\! N$ unique scene assets in GPU memory and shares these assets across subsets of the $N$ environments in a batch. To enable experience collection across a diverse set of environments, the renderer continuously updates the set of $K$ in-memory scene assets using asynchronous transfers that overlap rollout generation and learning.

SPL vs. wall-clock time (agents) on a RTX 3090 over 48 hours (time required to reach 2.5 billion samples with ). exceeds $80\%$ SPL in 10 hours and achieves a significantly higher SPL than the baselines. SPL vs. wall-clock time (training agents over 2.5 billion samples on 8 Tesla V100s) for various batch sizes ($N$). $N$=256 finishes after 2$\times$ the wall-clock time as $N$=1024, but both achieve statistically similar SPL.

's validation set SPL for vs. number of training samples across a range of batch sizes. This graph shows that sample efficiency slightly decreases with larger batch sizes (with the exception of $N$=512 vs. $N$=1024, where $N$=1024 exhibits better validation score). Ultimately, the difference in converged performance is less than 1% SPL between different batch sizes. Although $N$=256 converges the fastest in terms of training samples needed, {fig:batchspl} shows that $N$=256 performs poorly in terms of SPL achieved per unit of training time. Frames per second achieved by the standalone renderer on a RTX 3090 across a range of resolutions and batch sizes for a sensor on the Gibson dataset. Performance saturates at a batch size of 512. For lower batch sizes, increasing resolution has a minimal performance impact, because the GPU still isn't fully utilized. As resolution increases with larger batches, the relative decrease in performance from higher resolution increases.

The effect of the Lamb optimizer versus the baseline Adam optimizer on sample efficiency while training a sensor driven agent. Lamb maintains a consistent lead in terms of SPL throughout training, especially in the first half of training.

We run five trials of AP between $\mathcal A$ and $\mathcal B^{\text{lb}}$ with random initializations, where $N = 10$ and $R = 10$. For each trial, we plot the ratios $d(a_{k+1},E)/d(a_k,E)$, where $E =\mathcal A \cap \mathcal B^{\text{lb}}$ is the optimal set. The red line shows the theoretical lower bound of $1 - \tfrac{1}{R}(1-\cos(\tfrac{2\pi}{N}))$ on the worst-case rate of convergence.

Relative frequency errors $\tilde{e}_\infty (G, G_d^k)$, $\tilde{e}_\infty(G, P_\infty(G_d^k))$ versus the order $k$. A star marker indicates an unstable model.

Frequency-domain responses (up to the Nyquist frequency) of the continuous-time model $G$ and its discretised counterparts $G_d^{zoh}$, $G_d^{ii}$ and $G_d$.

Impulse responses of $G$ and its discretised counterparts $G_d^{zoh}$ (top), $G_d^{ii}$ (middle) and $G_d$ (bottom).

Time-domain error $y(t) - y_s(t)$ associated with the impulse responses of $G_d^{zoh}$, $G_d^{ii}$ and $G_d$.

Relative frequency errors $\tilde{e}_\infty (G, G_d^k)$, $\tilde{e}_\infty(G, P_\infty(G_d^k))$ versus the order $k$ for the TDS case (<ref>).

Frequency-domain responses of the time-delay system (<ref>) and its discretised counterparts.

Step responses of the time-delay system (<ref>) and its discretised counterparts $G_d^{tus}$ (top) and $G_d$ (bottom).

MiniCrafter environment: Example state from the environment. The water represents the traps, the $3$ resource types wood, iron, coal are present and a single crafting table.

Comparison of different SF pre-training experiments. HTR means Hindsight Task Replacement, TR means Task Replacement, $n$ and $1$ mean whether we used as many policies as features or a single one. The plots show the cumulative task competition on each task during training: Collect wood, Collect iron, Collect string and Collect table.

Pre-training experiments. Comparison of the best SF (Task Replacement with $n$ policies) against baselines. The plots show the cumulative task completion on each task during training: Collect wood, Collect iron, Collect string and Collect table.

Left side shows the performance of the best SF (standard SF with n policies in this case) compared to the baselines on the task "one_item". Middle shows the comparison of all SF training methods. Right side shows the comparison of all SF and baselines on the task "two_item".

Experiments on the "random" and "random_pen" targets. The plots show the running mean reward and the standard error (shaded area) during training. From left to right, the first and the third plots show the comparison of the best SF agent against the baseline agents, while the second and the final plots show the comparison of the SF agents with the different training methods used in this paper.

Left side shows the comparison of all SF training settings on the task "craft_staff". The other plots show all SFs compared to the baselines on all the crafting tasks.

Adversarial body shape search

Averages of the average cumulative rewards with attack strength $\epsilon=0.0,0.0.1,0.05,0.1$ for Walker2d-2d

Averages of the average cumulative rewards with attack strength $\epsilon=0.0,0.0.1,0.05,0.1$ for Ant-2d

Averages of the average cumulative rewards with attack strength $\epsilon=0.0,0.0.1,0.05,0.1$ for Humanoid-2d

Adversarial body shapes with length and thickness perturbations for Walker2d-v2.

Adversarial body shapes with length and thickness perturbations for Ant-v2.

Adversarial body shapes with length and thickness perturbations for Humanoid-v2.

{Coupling strength between dynamical quarks and the gluonic sector quantified through the center symmetry breaking $\Delta$ as a function of dimensionless curvature $R/T^2$.}

Illustration of the process of mapping the Lorentzian generators to the Euclidean generators. We start with quantizing the theory with respect to constant $P^0-K^0$ slices on the Lorentzian plane. Those are mapped to constant time slices on the Lorentzian cylinder. The Wick rotation maps those to constant time slices on the Euclidean cylinder. Finally we map the Euclidean cylinder to the Euclidean plane via a radial map.

Illustration of two nearby timelike geodesics in $\mathrm{AdS}_3$ (blue, red) corresponding to two boundary circuits and the minimal (green) and maximal (brown) perpendicular distance between them. The infinitesimal variation was exaggerated to improve the visualization.

Illustration of the test statistic and the calculation of the $p$-value of the null hypothesis.

$P$-value (color bars; in unit of %) of the null hypothesis (i.e. the multi-band light curves are consistent with Doppler boost plus DRW model) as a function of the DRW parameters $\sigma$ and $\tau$, with $A_{\rm DB}$ = 1 (top left), 1.5 (top right), 2.17 (middle left), 3 (middle right), 4 (bottom left) and 5 (bottom right) and UV/optical noise ratio $\sigma_{\rm UV}/\sigma_{\rm opt}=1$, considering the {Swift} $V$- and $M2$-bands. The black solid line represents the 5% $p$-value threshold, which separates models that are rejected (red) from those passing the test (blue). The star corresponds to the average $\sigma$ and $\tau$ for quasars with properties similar to PG1302-102 from <cit.>.

$P$-value (color bars; in unit of %) as a function $\sigma$ and $\tau$, for $A_{\rm DB}=2.17$ and with $r_{\rm noise}=\sigma_{\rm UV}/\sigma_{\rm opt}=1$ (top left), 2 (top right), 3 (bottom left) and 4 (bottom right), again considering the {Swift} $V$- and $M2$-bands.

UV spectra from HST (black) and GALEX (grey and light grey). The blue lines show power-law fits to the continuum. The transmission curves of the {Swift}/UVOT (and GALEX NUV) filters are shown to delineate the wavelength coverage of each band - the transmission curves are for illustrative purposes and are not measured in flux units ($y$-axis). The shaded band indicates the wavelength range in which the near-UV spectral slope was estimated in <cit.>.

Optical spectra taken with DBSP at Palomar and LRIS at Keck.

$P$-value (color bars; in unit of %) as a function of DRW parameters $\sigma$ and $\tau$ for six combinations of independent UV ($M2$ and $W1$) and optical ($V$ and $B$) bands, where $A_{\rm DB}$ = 2.17 in panels (a)-(d) and $A_{\rm DB}$ = 1 in (e) and (f), and UV/optical noise ratio $r_{\rm noise}=1$.

Comparison between ASAS-SN and {Swift} $V$-band photometry. The ASAS-SN data points are identical in both panels and are shown in light grey, except for the points taken closest to the time of the nine {Swift} observations analysed in this paper, which are shown in black. The red data points show {Swift} photometry from co-added images (top panel) and from the longest-exposure single image (bottom panel). The co-added {Swift} data are in better agreement with ASAS-SN.

Optical $V$-band (top panel) and near-UV $M2$-band (bottom panel) light curve of PG1301-102, similar to Figure <ref>. The prediction of the Doppler model with parameters from is shown with a solid light-blue line. The dashed dark-blue line represents a sinusoidal fit of the extended optical light curve, with period P=2095 d longer than in , and the corresponding Doppler boost prediction for the UV data with $A_{\rm DB}=2.17$ in the bottom panel.

$P$-values inferred when the fiducial model is fit to three different hypothetical future datasets with extended baselines. The red triangle denotes the original 9 {Swift} points. The dark blue curve is inferred from mock data generated in the fiducial model itself. The light blue curve assumes that the true light curve has a larger Doppler amplitude ratio ($A_{DB}=3$). The dark red curve corresponds to future mock data consisting of pure DRW variability.

Top Panel: Optical light curve of PG1302-102 with data from <cit.> in black (CRTS+LINEAR and other archival observations), data from ASAS-SN in grey, and purple squares/red diamonds for {Swift} $B$/$V$-band observations. Bottom panel: Near-UV light curve, with black circles and triangles for {GALEX} and {HST} observations from , purple squares and red diamonds for {Swift} $W1$ and $M2$-band observations, respectively. The sinusoidal Doppler boost model from is also shown in light blue.

$P$-values (color bars; in unit of %) of three pairs of independent combinations of bands {vs} the DRW parameters, assuming Doppler amplitude $A_{\rm DB}$ = 1 for the combinations of bands in panel (a) and $A_{\rm DB}$ = 2.17 in panels (b) and (C), and UV/optical noise ratio $r_{\rm noise}=1$.

Example hypothetical light curves over upcoming $\sim$10 years, for each of the three cases explained in <ref>; (1) the fiducial DB and DRW model with $A_{\rm DB}$=2.17 dominates the variability of PG1302 (left); (2) DB + DRW model has higher Doppler amplitude ratio, $A_{\rm DB}$=3 (middle); and (3) DB is absent from the system and DRW model dominates the true variability (right). In all three scenarios, the DRW amplitude ratio is fixed at 1 (i.e. $r_{\rm noise}$=1). The black points with error bars are the existing 9 {Swift} observations in optical (top panels) and UV (bottom panels). The red and dark blue points with error bars are the hypothetical optical and UV data, respectively.

Unimodal representation of a single modality can be either effective or not. The effectiveness of different unimodal representations from the same sample also varies. To empower the interaction between modalities, our proposed method aligns the unimodal representation to the effective modality sample-wise and makes full use of the effective unimodal representation under the supervision of the unimodal prediction (F and T represent correct and incorrect predictions, respectively).

The relationship comparison between two modalities in training mini-batch of (a) unsupervised MMC, (b) supervised MMC and (c) UniS-MMC.

Unimodal representation distribution of the first 10 categories of the N24News test set across different methods: (a) aggregation-based method, (b) unsupervised multimodal method, (c) supervised contrastive method and (d) unimodality-supervised method.

Multimodal representation distribution of the first 10 categories of the N24News test set across different methods: (a) aggregation-based method, (b) unsupervised multimodal method, (c) supervised contrastive method and (d) unimodality-supervised method.

As the training progresses, the change of the proportion of both wrong (left), both correct (right) unimodal predictions of the validation set (N24News): the complete method (UniS-MMC), remove negative pair (w.o. C_{Neg}), remove semi-positive pair (w.o. C_{Semi}) and remove both (w.o. C_{Neg},C_{Semi}).

Consistency comparison of unimodal prediction between MT-MML and the UniS-MMC.

The framework for our proposed UniS-MMC.

Loss function for L1 estimator Loss function for L2 estimator

Estimation for three-parameter model.

Estimation for one-parameter model.

(a) A DL model's screening accuracy in different label uncertainty groups; (b) GON severity distributions on different label uncertainty. Both x-axes are label uncertainty scores.

Image examples with different GON severity levels and empirical uncertainty scores from the GON dataset.

Overview of the training phase and the inference phase of the uncertainty-guided multi-stream screening model.

Computed margins $\lambda^j_t$, effective disturbances $\hat{\delta}^j_t$ and disturbance $w^j_t$ for every node $x^j$. $\mu_t = \sum\limits_{k} \left\| \Delta_t(k)\right\|_1$ is computed with the real system and the controller at $t$.

State and input trajectories for closed loop simulation of controller {algo:DLAR} with perfect parameter knowledge $\alpha$.

Left: Overlay of projections of $\Pt{t}^j$ onto different coordinates for different time steps. Each row corresponds to a different subsystem $x_1$, $x_3$, $x_5$ (top to bottom). Shading indicates time of computation with shades lightening as simulation time passes. Right: state and input trajectories of closed loop simulation with {algo:DLAR} and uncertainties on $\alpha$ described in {sec:introex}.

(A) The initial bipartite network structure. Each ego is connected to $2$ alters out of $6$. (B) Protocol for each of the $5$ rounds. In turn-1, an ego (blue) generates ideas independently. In turn-2, they view the ideas of the two alters (red) they are following, and submit any new ideas inspired by the stimuli. Then, they rate the ideas of all $6$ alters, and update which two alters to follow in the next round. (C) The $4$ study conditions. The top, middle and bottom tiers of alters, as recorded from C1, are respectively shown in teal, mustard, and orange bars. (D) Non-redundant idea counts across conditions. Whiskers denote $95\%$ C.I.

Examples of different TST. Including sentiment style transfer (negative $\leftrightarrow$ positive), formality style transfer (informal $\leftrightarrow$ formal), code-switch style transfer (single language $\leftrightarrow$ code-switch sentence).

Comparison of style transfer outputs of our models and style transformer in three transfer tasks. Our models are stage II models in Table <ref>. Translations for code-switching sentences are shown in parenthesis.

Illustration of style weight $w$ vs. Acc, PPL and BLEU in sentiment, formality and code-switching transfer tasks. Note there is a trade-off between Acc and PPL/BLEU. With increasing of $w$, Acc will increase while BLEU drops down and PPL increases. sentiment transfer

Illustration of style weight $w$ vs. Acc, PPL and BLEU in sentiment, formality and code-switching transfer tasks. Note there is a trade-off between Acc and PPL/BLEU. With increasing of $w$, Acc will increase while BLEU drops down and PPL increases. formality transfer

Illustration of style weight $w$ vs. Acc, PPL and BLEU in sentiment, formality and code-switching transfer tasks. Note there is a trade-off between Acc and PPL/BLEU. With increasing of $w$, Acc will increase while BLEU drops down and PPL increases. code-switching transfer

Illustration of the mean attention weights of token `$<$s$>$' from all heads at the final layer in three TST tasks. Higher importance scores are assigned to pivot words, which are depicted as deeper lines in the figures. sentiment transfer

Illustration of the mean attention weights of token `$<$s$>$' from all heads at the final layer in three TST tasks. Higher importance scores are assigned to pivot words, which are depicted as deeper lines in the figures. formality transfer

Illustration of the mean attention weights of token `$<$s$>$' from all heads at the final layer in three TST tasks. Higher importance scores are assigned to pivot words, which are depicted as deeper lines in the figures. code-switching transfer

A single-cell MRC-WPT network with $K$ receivers.

Outage probability in the strong coupling region; Markers and dashed lines correspond to simulations and analytical results, respectively.

Outage probability in the loosely coupling region; Markers and dashed lines correspond to simulations and analytical results, respectively.

Harvested power at each receiver with different loads.

Atomic levels and phase matching in $\Lambda$-scheme Raman scattering induced by classical laser field $\mathcal{E}_{\mathrm{W}}$ and $\mathcal{E}_{\mathrm{R}}$. (a) In the spontaneous write-in process Stokes photons ($\protect\ad_{\mathrm{WS}}$ - mode) and spin-wave excitation ($\protect\bd$ - mode) are created pairwise. (b) Four-wave mixing in readout consists in simultaneous anti-Stokes and Stokes scattering into modes $\protect\ad_{\mathrm{RA}}$ and $\protect\ad_{\mathrm{RS}}$. (c) Phase matching condition or momentum conservation dictates the wave vectors of single photons coupled to a spin-wave excitation with a certain wave vector $\mathbf{K}_{b}$. Write beam with $\mathbf{k}_{\mathrm{W}}$ wave vector is scattered as Stokes photon with $\mathbf{k}_{\mathrm{WS}}$ wave vector, while read beam with $\mathbf{k}_{\mathrm{R}}$ wave vector either scatters Stokes or couples anti-Stokes of the respective wave vectors $\mathbf{k}_{\mathrm{RS}}$, $\mathbf{k}_{\mathrm{RA}}$.

(a) to (c) - temporal evolution of gains $G_{i}$ and spontaneous noises $S_{i}$ building up the light fields. Data plotted for different values of coupling coefficients $\chi$ and $\xi$ correspond to detuning parameters $\Delta_{R}$= 0.3, 0.6 and 0.8. (d) Integrated readout gains $\bar{G}_{i}$ and spontaneous noises $\bar{S}_{i}$ versus detuning $\Delta_{R}$.