Search

The Fr\'echet Video Distance (FVD) is a widely adopted metric for evaluating video generation distribution quality. However, its effectiveness relies on critical assumptions. Our analysis reveals three significant limitations: (1) the non-Gaussianity of the Inflated 3D Convnet (I3D) feature space; (2) the insensitivity of I3D features to temporal distortions; (3) the impractical sample sizes required for reliable estimation. These findings undermine FVD's reliability and show that FVD falls short as a standalone metric for video generation evaluation. After extensive analysis of a wide range of metrics and backbone architectures, we propose JEDi, the JEPA Embedding Distance, based on features derived from a Joint Embedding Predictive Architecture, measured using Maximum Mean Discrepancy with polynomial kernel. Our experiments on multiple open-source datasets show clear evidence that it is a superior alternative to the widely used FVD metric, requiring only 16% of the samples to reach its steady value, while increasing alignment with human evaluation by 34%, on average.

The fluctuations produced during cosmic inflation may exhibit non-Gaussian characteristics that are imprinted in the large-scale structure of the Universe. This non-Gaussian imprint is an ultra-large scale signal that can be detected using the power spectrum. We focus on the local-type non-Gaussianity $f_{\rm{NL}}$ and employ a multi-tracer analysis that combines different probes in order to mitigate cosmic variance and maximize the non-Gaussian signal. In our previous paper, we showed that combining spectroscopic galaxy surveys with 21cm intensity mapping surveys in interferometer mode could lead to a $\sim 20-30\%$ improvement in the precision on this non-Gaussian signal. Here we combine the same 21 cm experiments, including also single-dish surveys, with photometric galaxy surveys. The 21 cm single-dish surveys are based on MeerKAT and SKAO and the interferometric surveys are alike to HIRAX and PUMA. We implement foreground-avoidance filters and utilize models for the 21 cm thermal noise associated with single-dish and interferometer modes. The photometric galaxy surveys are similar to the DES and LSST. Our multi-tracer Fisher forecasts show a better precision for the combination of the photometric galaxy surveys and 21 cm interferometric surveys than with the 21 cm single-dish surveys -leading to at most an improvement of $23\%$ in the former case and $16\%$ in the latter case. Furthermore, we examine the impact of varying the foreground filter parameter, redshift range and sky area on the derived constraint. We find that the $f_{\rm{NL}}$ constraint is highly sensitive to both the redshift range and sky area. The foreground filter parameter shows negligible effect., Comment: 14 pages, 8 figures

We discuss the magnetic ground state and properties of a frustrated two-dimensional classical Heisenberg model of interacting hexagonal clusters of spins. The energy of the ground states is found exactly for arbitrary values of $J_1$ (intra-cluster couplings) and $J_2$ (inter-cluster couplings). Our main results concern a frustrated region of the phase diagram, where we show that the set of ground states has a degeneracy larger than that due to global rotation symmetry. Furthermore, the ground state manifold does not have a fixed total magnetization~: there is a range of allowed values. At finite temperature, our Monte-Carlo simulations show that the entropy selects the most probable value of the total magnetization, while the histogram of the Monte-Carlo time series is non-trivial. This model is a first step towards modelling properties of a class of frustrated magnetic structures composed of coupled spin clusters., Comment: 6 pages, 7 figures, submitted to Phys Rev B Letters

Generating novel molecules is challenging, with most representations leading to generative models producing many invalid molecules. Spanning Tree-based Graph Generation (STGG) is a promising approach to ensure the generation of valid molecules, outperforming state-of-the-art SMILES and graph diffusion models for unconditional generation. In the real world, we want to be able to generate molecules conditional on one or multiple desired properties rather than unconditionally. Thus, in this work, we extend STGG to multi-property-conditional generation. Our approach, STGG+, incorporates a modern Transformer architecture, random masking of properties during training (enabling conditioning on any subset of properties and classifier-free guidance), an auxiliary property-prediction loss (allowing the model to self-criticize molecules and select the best ones), and other improvements. We show that STGG+ achieves state-of-the-art performance on in-distribution and out-of-distribution conditional generation, and reward maximization., Comment: Code: https://github.com/SamsungSAILMontreal/AnyMolGenCritic

Galaxy surveys contain information on the largest scales via wide-angle and relativistic contributions. By combining two different galaxy populations, we can suppress the strong cosmic variance on ultra-large scales and thus enhance the detectability of the signals. The relativistic Doppler and Sachs-Wolfe effects are of a similar magnitude to the leading wide-angle corrections, so that it is important to treat them together, especially since they can partially cancel. The power spectra depend on the choice of line of sight for each galaxy pair and we present results for a general line of sight. Then we estimate the detection significance of the auto- and cross-power spectra for a variety of cases. We use two futuristic galaxy samples based on a `beyond-DESI' survey and a SKA Phase 2 survey, covering 15,000\,deg$^2$ up to $z=1$. We find a detection significance for the total relativistic wide-angle effects that ranges from $\sim 5\sigma$ to $>15\sigma$, depending on the line-of-sight configuration., Comment: 23 pages, 9 figures, 'version accepted by JCAP'

With recent advances in video prediction, controllable video generation has been attracting more attention. Generating high fidelity videos according to simple and flexible conditioning is of particular interest. To this end, we propose a controllable video generation model using pixel level renderings of 2D or 3D bounding boxes as conditioning. In addition, we also create a bounding box predictor that, given the initial and ending frames' bounding boxes, can predict up to 15 bounding boxes per frame for all the frames in a 25-frame clip. We perform experiments across 3 well-known AV video datasets: KITTI, Virtual-KITTI 2 and BDD100k.

A good initialization of deep learning models is essential since it can help them converge better and faster. However, pretraining large models is unaffordable for many researchers, which makes a desired prediction for initial parameters more necessary nowadays. Graph HyperNetworks (GHNs), one approach to predicting model parameters, have recently shown strong performance in initializing large vision models. Unfortunately, predicting parameters of very wide networks relies on copying small chunks of parameters multiple times and requires an extremely large number of parameters to support full prediction, which greatly hinders its adoption in practice. To address this limitation, we propose LoGAH (Low-rank GrAph Hypernetworks), a GHN with a low-rank parameter decoder that expands to significantly wider networks without requiring as excessive increase of parameters as in previous attempts. LoGAH allows us to predict the parameters of 774-million large neural networks in a memory-efficient manner. We show that vision and language models (i.e., ViT and GPT-2) initialized with LoGAH achieve better performance than those initialized randomly or using existing hypernetworks. Furthermore, we show promising transfer learning results w.r.t. training LoGAH on small datasets and using the predicted parameters to initialize for larger tasks. We provide the codes in https://github.com/Blackzxy/LoGAH ., Comment: 16 pages

Despite substantial progress in machine learning for scientific discovery in recent years, truly de novo design of small molecules which exhibit a property of interest remains a significant challenge. We introduce LambdaZero, a generative active learning approach to search for synthesizable molecules. Powered by deep reinforcement learning, LambdaZero learns to search over the vast space of molecules to discover candidates with a desired property. We apply LambdaZero with molecular docking to design novel small molecules that inhibit the enzyme soluble Epoxide Hydrolase 2 (sEH), while enforcing constraints on synthesizability and drug-likeliness. LambdaZero provides an exponential speedup in terms of the number of calls to the expensive molecular docking oracle, and LambdaZero de novo designed molecules reach docking scores that would otherwise require the virtual screening of a hundred billion molecules. Importantly, LambdaZero discovers novel scaffolds of synthesizable, drug-like inhibitors for sEH. In in vitro experimental validation, a series of ligands from a generated quinazoline-based scaffold were synthesized, and the lead inhibitor N-(4,6-di(pyrrolidin-1-yl)quinazolin-2-yl)-N-methylbenzamide (UM0152893) displayed sub-micromolar enzyme inhibition of sEH.

The fluctuations generated by Inflation are nearly Gaussian in the simplest models, but may be non-Gaussian in more complex models, potentially leading to signatures in the late Universe. In particular, local type primordial non-Gaussianity induces scale-dependent bias in tracers of the matter distribution. This non-Gaussian imprint in the tracer power spectrum survives at late times on ultra-large scales where nonlinearity is negligible. In order to combat the problem of growing cosmic variance on these scales, we use a multi-tracer analysis that combines different tracers to maximise any primordial signal. Previous work has investigated the combination of a spectroscopic galaxy survey with a 21cm intensity mapping survey in single-dish mode. We extend this work by considering instead the case where the 21cm intensity mapping survey is optimised for interferometer mode. As examples, we use two multi-tracer pairs of surveys: one at high redshift ($1 \le z \le 2$) and one at very high redshift ($2 \le z \le 5$). The 21cm surveys are idealised surveys based on HIRAX and PUMA. We implement foreground-avoidance filters and use detailed models of the interferometer thermal noise. The galaxy surveys are idealised surveys based on Euclid and MegaMapper. Via a simple Fisher forecast, we illustrate the potential of the multi-tracer. Our results show a $\sim 20 - 30\%$ improvement in precision on local primordial non-Gaussianity from the multi-tracer. Furthermore, we investigate the effects on constraints of varying the parameter of non-Gaussian galaxy assembly bias and of varying the parameters of the intensity mapping foreground filters. We find that the non-Gaussian galaxy assembly bias parameter causes a greater change in the constraints on local primordial non-Gaussianity than the foreground filter parameters., Comment: 15 pages, 8 figures. Minor improvements and corrections. Version accepted by EPJC

In the pursuit of understanding the large-scale structure of the Universe, the synergy between complementary cosmological surveys has proven to be a powerful tool. Using multiple tracers of the large-scale structure can significantly improve the constraints on cosmological parameters. We explore the potential of combining the Square Kilometre Array Observatory (SKAO) and the Dark Energy Spectroscopic Instrument (DESI) spectroscopic surveys to enhance precision on the growth rate of cosmic structures. We employ a multitracer Fisher analysis to estimate precision on the growth rate when using pairs of mock surveys that are based on SKAO and DESI specifications. The pairs are at both low and high redshifts. For SKA-MID, we use the HI galaxy and the HI intensity mapping samples. In order to avoid the complexities and uncertainties at small scales, we confine the analysis to scales where linear perturbations are reliable. The consequent loss of signal in each individual survey is mitigated by the gains from the multi-tracer. After marginalising over cosmological and nuisance parameters, we find a significant improvement in the precision on the growth rate., Comment: 12 pages, 8 figures

Tabular data is hard to acquire and is subject to missing values. This paper introduces a novel approach for generating and imputing mixed-type (continuous and categorical) tabular data utilizing score-based diffusion and conditional flow matching. In contrast to prior methods that rely on neural networks to learn the score function or the vector field, we adopt XGBoost, a widely used Gradient-Boosted Tree (GBT) technique. To test our method, we build one of the most extensive benchmarks for tabular data generation and imputation, containing 27 diverse datasets and 9 metrics. Through empirical evaluation across the benchmark, we demonstrate that our approach outperforms deep-learning generation methods in data generation tasks and remains competitive in data imputation. Notably, it can be trained in parallel using CPUs without requiring a GPU. Our Python and R code is available at https://github.com/SamsungSAILMontreal/ForestDiffusion., Comment: Code: https://github.com/SamsungSAILMontreal/ForestDiffusion

We theoretically and numerically investigate the spin fluctuations induced in a thermal atomic ensemble by an external fluctuating uniaxial magnetic field, in the context of a standard spin noise spectroscopy (SNS) experiment. We show that additional spin noise is excited, which dramatically depends on the magnetic noise variance and bandwidth, as well as on the power of the probe light and its polarization direction. We develop an analytical perturbative model proving that this spin noise first emerges from the residual optical pumping in the medium, which is then converted into spin fluctuations by the magnetic noise and eventually detected using SNS. The system studied is a spin-1 system, which thus shows both Faraday rotation and ellipticity noises induced by the random magnetic fluctuations. The analytical model gives results in perfect agreement with the numerical simulations, with potential applications in future experimental characterization of stray field properties and their influence on spin dynamics., Comment: 16 pages, 10 figures, submitted to New Journal of Physics

Diffusion Models (DMs) are powerful generative models that add Gaussian noise to the data and learn to remove it. We wanted to determine which noise distribution (Gaussian or non-Gaussian) led to better generated data in DMs. Since DMs do not work by design with non-Gaussian noise, we built a framework that allows reversing a diffusion process with non-Gaussian location-scale noise. We use that framework to show that the Gaussian distribution performs the best over a wide range of other distributions (Laplace, Uniform, t, Generalized-Gaussian).

Ensemble methods combine the predictions of multiple models to improve performance, but they require significantly higher computation costs at inference time. To avoid these costs, multiple neural networks can be combined into one by averaging their weights. However, this usually performs significantly worse than ensembling. Weight averaging is only beneficial when different enough to benefit from combining them, but similar enough to average well. Based on this idea, we propose PopulAtion Parameter Averaging (PAPA): a method that combines the generality of ensembling with the efficiency of weight averaging. PAPA leverages a population of diverse models (trained on different data orders, augmentations, and regularizations) while slowly pushing the weights of the networks toward the population average of the weights. We also propose PAPA variants (PAPA-all, and PAPA-2) that average weights rarely rather than continuously; all methods increase generalization, but PAPA tends to perform best. PAPA reduces the performance gap between averaging and ensembling, increasing the average accuracy of a population of models by up to 0.8% on CIFAR-10, 1.9% on CIFAR-100, and 1.6% on ImageNet when compared to training independent (non-averaged) models., Comment: Blog post: https://ajolicoeur.wordpress.com/papa/, Code: https://github.com/SamsungSAILMontreal/PAPA, TMLR journal publication: https://openreview.net/forum?id=cPDVjsOytS

Surveys of the matter distribution contain `fossil' information on possible non-Gaussianity that is generated in the primordial Universe. This primordial signal survives only on the largest scales where cosmic variance is strongest. By combining different surveys in a multi-tracer approach, we can suppress the cosmic variance and significantly improve the precision on the level of primordial non-Gaussianity. We consider a combination of an optical galaxy survey, like the recently initiated DESI survey, together with a new and very different type of survey, a 21 cm intensity mapping survey, like the upcoming SKAO survey. A Fisher forecast of the precision on the local primordial non-Gaussianity parameter $f_{\mathrm{NL}}$, shows that this multi-tracer combination, together with non-overlap single-tracer information, can deliver precision comparable to that from the CMB. Taking account of the largest systematic, i.e. foreground contamination in intensity mapping, we find that $\sigma(f_{\mathrm{NL}}) \sim 4$., Comment: 10+5 pages, 4 figures, "Minor corrections, main results unchanged. Version accepted by EPJC"

We propose a novel regularizer for supervised learning called Conditioning on Noisy Targets (CNT). This approach consists in conditioning the model on a noisy version of the target(s) (e.g., actions in imitation learning or labels in classification) at a random noise level (from small to large noise). At inference time, since we do not know the target, we run the network with only noise in place of the noisy target. CNT provides hints through the noisy label (with less noise, we can more easily infer the true target). This give two main benefits: 1) the top-down feedback allows the model to focus on simpler and more digestible sub-problems and 2) rather than learning to solve the task from scratch, the model will first learn to master easy examples (with less noise), while slowly progressing toward harder examples (with more noise).

We study the fractional quantum Hall effect in the central Landau level of bilayer graphene. By tuning the external applied magnetic field and the electric bias between the two layers one can access a regime where there is a degeneracy between Landau levels with orbital characters corresponding to N=0 and N=1 Galilean Landau levels. While the Laughlin state is generically the ground state for filling $\nu=1/3$ we find that it can be destroyed and replaced by a Wigner crystal at the same filling factor by tuning the bias and applied field. This competition does not take place at $\nu=2/3$ where the incompressible ground state remains stable. The possibility of electrically inducing the Wigner crystal state opens a new range of studies of this state of matter., Comment: 11 pages and 11 figures, v2: references corrected

Video prediction is a challenging task. The quality of video frames from current state-of-the-art (SOTA) generative models tends to be poor and generalization beyond the training data is difficult. Furthermore, existing prediction frameworks are typically not capable of simultaneously handling other video-related tasks such as unconditional generation or interpolation. In this work, we devise a general-purpose framework called Masked Conditional Video Diffusion (MCVD) for all of these video synthesis tasks using a probabilistic conditional score-based denoising diffusion model, conditioned on past and/or future frames. We train the model in a manner where we randomly and independently mask all the past frames or all the future frames. This novel but straightforward setup allows us to train a single model that is capable of executing a broad range of video tasks, specifically: future/past prediction -- when only future/past frames are masked; unconditional generation -- when both past and future frames are masked; and interpolation -- when neither past nor future frames are masked. Our experiments show that this approach can generate high-quality frames for diverse types of videos. Our MCVD models are built from simple non-recurrent 2D-convolutional architectures, conditioning on blocks of frames and generating blocks of frames. We generate videos of arbitrary lengths autoregressively in a block-wise manner. Our approach yields SOTA results across standard video prediction and interpolation benchmarks, with computation times for training models measured in 1-12 days using $\le$ 4 GPUs. Project page: https://mask-cond-video-diffusion.github.io ; Code : https://github.com/voletiv/mcvd-pytorch, Comment: NeurIPS 2022 ; 10 pages, 4 figures, 7 tables

We study spin and valley ordering in the quantum Hall fractions in monolayer graphene at Landau level filling factors $\nu_G=-2+n/3$ $(n=2,4,5)$. We use exact diagonalizations on the spherical as well as toroidal geometry by taking into account the effect of realistic anisotropies that break the spin/valley symmetry of the pure Coulomb interaction. We also use a variational method based on eigenstates of the fully $SU(4)$ symmetric limit. For all the fractions we study there are two-component states for which the competing phases are generalizations of those occurring at neutrality $\nu_G=0$. They are ferromagnetic, antiferromagnetic, charge-density wave and K\'ekul\'e phases, depending on the values of Ising or XY anisotropies in valley space. The varying spin-valley content of the states leads to ground state quantum numbers that are different from the $\nu_G=0$ case. For filling factor $\nu_G=-2+5/3$ there is a parent state in the $SU(4)$ limit which has a flavor content $(1,1/3,1/3,0)$ where the two components that are one-third filled form a two-component singlet. The addition of anisotropies leads to the formation of new states that have no counterpart at $\nu_G=0$. While some of them are predicted by the variational approach, we find notably that negative Ising-like valley anisotropy leads to the formation of a state which is a singlet in both spin and valley space and lies beyond the reach of the variational method. Also fully spin polarized two-component states at $\nu=-2+4/3$ and $\nu=-2+5/3$ display an emergent $SU(2)$ valley symmetry because they do not feel point-contact anisotropies. We discuss implications for current experiments concerning possible spin transitions., Comment: 16 pages, 7 figures

Measurements of galaxy clustering in upcoming surveys such as those planned for the Euclid and Roman satellites, and the SKA Observatory, will be sensitive to distortions from lensing magnification and Doppler effects, beyond the standard redshift-space distortions. The amplitude of these contributions depends sensitively on magnification bias and evolution bias in the galaxy number density. Magnification bias quantifies the change in the observed number of galaxies gained or lost by lensing magnification, while evolution bias quantifies the physical change in the galaxy number density relative to the conserved case. These biases are given by derivatives of the number density, and consequently are very sensitive to the form of the luminosity function. We give a careful derivation of the magnification and evolution biases, clarifying a number of results in the literature. We then examine the biases for a variety of surveys, encompassing galaxy surveys and line intensity mapping at radio and optical/near-infrared wavelengths., Comment: Typo corrected in equation (A.10)

Score-based (denoising diffusion) generative models have recently gained a lot of success in generating realistic and diverse data. These approaches define a forward diffusion process for transforming data to noise and generate data by reversing it (thereby going from noise to data). Unfortunately, current score-based models generate data very slowly due to the sheer number of score network evaluations required by numerical SDE solvers. In this work, we aim to accelerate this process by devising a more efficient SDE solver. Existing approaches rely on the Euler-Maruyama (EM) solver, which uses a fixed step size. We found that naively replacing it with other SDE solvers fares poorly - they either result in low-quality samples or become slower than EM. To get around this issue, we carefully devise an SDE solver with adaptive step sizes tailored to score-based generative models piece by piece. Our solver requires only two score function evaluations, rarely rejects samples, and leads to high-quality samples. Our approach generates data 2 to 10 times faster than EM while achieving better or equal sample quality. For high-resolution images, our method leads to significantly higher quality samples than all other methods tested. Our SDE solver has the benefit of requiring no step size tuning., Comment: Code is available on https://github.com/AlexiaJM/score_sde_fast_sampling

Active fluid droplets surrounded by oil can spontaneously develop circulatory flows. However, the dynamics of the surrounding oil and their influence on the active fluid remain poorly understood. To investigate interactions between the active fluid and the passive oil across their interface, kinesin-driven microtubule-based active fluid droplets were immersed in oil and compressed into a cylinder-like shape. The droplet geometry supported intradroplet circulatory flows, but the circulation was suppressed when the thickness of the oil layer surrounding the droplet decreased. Experiments with tracers and network structure analyses and continuum models based on the dynamics of self-elongating rods demonstrated that the flow transition resulted from flow coupling across the interface between active fluid and oil, with a millimeter-scale coupling length. In addition, two novel millifluidic devices were developed that could trigger and suppress intradroplet circulatory flows in real time: one by changing the thickness of the surrounding oil layer and the other by locally deforming the droplet. This work highlights the role of interfacial dynamics in the active fluid droplet system and shows that circulatory flows within droplets can be affected by millimeter-scale flow coupling across the interface between the active fluid and the oil.

Magnons enable transferring a magnetic moment or spin over macroscopic distance. In quantum Hall ferromagnet, it has been predicted in the early 90s that spin and charges are entangled, meaning that any change of the spin texture modifies the charge distribution. As a direct consequence of this entanglement, magnons carry an electric dipole moment. Here we report the first evidence of the existence of this electric dipole moment in a graphene quantum Hall ferromagnet using a Mach-Zehnder interferometer as a quantum sensor. By propagating towards the interferometer through an insulating bulk, the magnon electric dipole moment modifies the Aharonov-Bohm flux through the interferometer, changing both its phase and its visibility. In particular, we relate the phase shift to the sign of this electric dipole moment, and the exponential loss of visibility to the flux of emitted magnons. Finally, we probe the emission energy threshold of the magnons close to filling factor v=1. Approaching v=0, we observe that the emission energy threshold diminishes towards zero, which might be linked to the existence of gapless mode in the canted-antiferromagnetic (CAF) phase at v=0. The detection and manipulation of magnons based on their electric dipole open the field for a new type of coherent magnon quantum circuits that will be electrostatically controlled., Comment: 11 pages, 4 figures