Search

Multimodal Large Language Models (MLLMs) have shown promising progress in understanding and analyzing video content. However, processing long videos remains a significant challenge constrained by LLM's context size. To address this limitation, we propose LongVU, a spatiotemporal adaptive compression mechanism thats reduces the number of video tokens while preserving visual details of long videos. Our idea is based on leveraging cross-modal query and inter-frame dependencies to adaptively reduce temporal and spatial redundancy in videos. Specifically, we leverage DINOv2 features to remove redundant frames that exhibit high similarity. Then we utilize text-guided cross-modal query for selective frame feature reduction. Further, we perform spatial token reduction across frames based on their temporal dependencies. Our adaptive compression strategy effectively processes a large number of frames with little visual information loss within given context length. Our LongVU consistently surpass existing methods across a variety of video understanding benchmarks, especially on hour-long video understanding tasks such as VideoMME and MLVU. Given a light-weight LLM, our LongVU also scales effectively into a smaller size with state-of-the-art video understanding performance., Comment: Project page: https://vision-cair.github.io/LongVU

Biological sequences encode fundamental instructions for the building blocks of life, in the form of DNA, RNA, and proteins. Modeling these sequences is key to understand disease mechanisms and is an active research area in computational biology. Recently, Large Language Models have shown great promise in solving certain biological tasks but current approaches are limited to a single sequence modality (DNA, RNA, or protein). Key problems in genomics intrinsically involve multiple modalities, but it remains unclear how to adapt general-purpose sequence models to those cases. In this work we propose a multi-modal model that connects DNA, RNA, and proteins by leveraging information from different pre-trained modality-specific encoders. We demonstrate its capabilities by applying it to the largely unsolved problem of predicting how multiple RNA transcript isoforms originate from the same gene (i.e. same DNA sequence) and map to different transcription expression levels across various human tissues. We show that our model, dubbed IsoFormer, is able to accurately predict differential transcript expression, outperforming existing methods and leveraging the use of multiple modalities. Our framework also achieves efficient transfer knowledge from the encoders pre-training as well as in between modalities. We open-source our model, paving the way for new multi-modal gene expression approaches.

Following the recent popularity of Large Language Models (LLMs), several attempts have been made to extend them to the visual domain. From having a visual assistant that could guide us through unfamiliar environments to generative models that produce images using only a high-level text description, the vision-language model (VLM) applications will significantly impact our relationship with technology. However, there are many challenges that need to be addressed to improve the reliability of those models. While language is discrete, vision evolves in a much higher dimensional space in which concepts cannot always be easily discretized. To better understand the mechanics behind mapping vision to language, we present this introduction to VLMs which we hope will help anyone who would like to enter the field. First, we introduce what VLMs are, how they work, and how to train them. Then, we present and discuss approaches to evaluate VLMs. Although this work primarily focuses on mapping images to language, we also discuss extending VLMs to videos.

Curation methods for massive vision-language datasets trade off between dataset size and quality. However, even the highest quality of available curated captions are far too short to capture the rich visual detail in an image. To show the value of dense and highly-aligned image-text pairs, we collect the Densely Captioned Images (DCI) dataset, containing 7805 natural images human-annotated with mask-aligned descriptions averaging above 1000 words each. With precise and reliable captions associated with specific parts of an image, we can evaluate vision-language models' (VLMs) understanding of image content with a novel task that matches each caption with its corresponding subcrop. As current models are often limited to 77 text tokens, we also introduce a summarized version (sDCI) in which each caption length is limited. We show that modern techniques that make progress on standard benchmarks do not correspond with significant improvement on our sDCI based benchmark. Lastly, we finetune CLIP using sDCI and show significant improvements over the baseline despite a small training set. By releasing the first human annotated dense image captioning dataset, we hope to enable the development of new benchmarks or fine-tuning recipes for the next generation of VLMs to come.

Constraints on the quantum decoherence of entangled $\gamma$ quanta at the mega-electron-volt scale, such as those produced following positron annihilation, have remained elusive for many decades. We present the first statistically and kinematically precise experimental data for triple Compton scattering of such entangled $\gamma$. An entanglement witness ($R$), relating to the enhancement of the azimuthal correlation between the final scattering planes, is obtained where one of the $\gamma$ underwent intermediate Compton scattering. The measured $R$, deconvolved from multiple scattering backgrounds, are found to exceed the classical limit for intermediate scatter angles up to $\sim$60$^{\circ}$ and diminish at larger angles. The data are consistent with predictions from a first quantum theory of entangled triple Compton scattering as well as a simple model based approach. The results are crucial to future study and utilisation of entangled mega-electron-volt $\gamma$ in fundamental physics and PET imaging., Comment: 5 pages, 4 figures

The framed standard model (FSM), constructed to explain, with some success, why there should be 3 and apparently only 3 generations of quarks and leptons in nature falling into a hierarchical mass and mixing pattern, suggests also, among other things, a scalar boson U, with mass around 17 MeV and small couplings to quarks and leptons, which might explain the g-2 anomaly reported in experiment. The U arises in FSM initially as a state in the predicted `hidden sector' with mass around 17 MeV, which mixes with the standard model (SM) Higgs $h_W$, acquiring thereby a coupling to quarks and leptons and a mass just below 17 MeV. The initial purpose of the present paper is to check whether this proposal is compatible with experiment on semileptonic decays of Ks and Bs where the U can also appear. The answer to this we find is affirmative, in that the contribution of U to new physics as calculated in the FSM remains within the experimental bounds, but only if $m_U$ lies within a narrow range just below the unmixed mass. As a result from this, one has an estimate $m_U \sim 15 - 17$ MeV for the mass of $U$, and from some further considerations the estimate $\Gamma_U \sim 0.02$ eV for its width, both of which may be useful for an eventual search for it in experiment. And, if found, it will be, for the FSM, not just the discovery of a predicted new particle, but the opening of a window into a whole ``hidden sector" containing at least some, perhaps ven the bulk, of the dark matter in the universe.

Current status quo in machine learning is to use static datasets of real images for training, which often come from long-tailed distributions. With the recent advances in generative models, researchers have started augmenting these static datasets with synthetic data, reporting moderate performance improvements on classification tasks. We hypothesize that these performance gains are limited by the lack of feedback from the classifier to the generative model, which would promote the usefulness of the generated samples to improve the classifier's performance. In this work, we introduce a framework for augmenting static datasets with useful synthetic samples, which leverages one-shot feedback from the classifier to drive the sampling of the generative model. In order for the framework to be effective, we find that the samples must be close to the support of the real data of the task at hand, and be sufficiently diverse. We validate three feedback criteria on a long-tailed dataset (ImageNet-LT) as well as a group-imbalanced dataset (NICO++). On ImageNet-LT, we achieve state-of-the-art results, with over 4 percent improvement on underrepresented classes while being twice efficient in terms of the number of generated synthetic samples. NICO++ also enjoys marked boosts of over 5 percent in worst group accuracy. With these results, our framework paves the path towards effectively leveraging state-of-the-art text-to-image models as data sources that can be queried to improve downstream applications.

Synthetic image datasets offer unmatched advantages for designing and evaluating deep neural networks: they make it possible to (i) render as many data samples as needed, (ii) precisely control each scene and yield granular ground truth labels (and captions), (iii) precisely control distribution shifts between training and testing to isolate variables of interest for sound experimentation. Despite such promise, the use of synthetic image data is still limited -- and often played down -- mainly due to their lack of realism. Most works therefore rely on datasets of real images, which have often been scraped from public images on the internet, and may have issues with regards to privacy, bias, and copyright, while offering little control over how objects precisely appear. In this work, we present a path to democratize the use of photorealistic synthetic data: we develop a new generation of interactive environments for representation learning research, that offer both controllability and realism. We use the Unreal Engine, a powerful game engine well known in the entertainment industry, to produce PUG (Photorealistic Unreal Graphics) environments and datasets for representation learning. In this paper, we demonstrate the potential of PUG to enable more rigorous evaluations of vision models.

Masked Image Modeling (MIM) is a promising self-supervised learning approach that enables learning from unlabeled images. Despite its recent success, learning good representations through MIM remains challenging because it requires predicting the right semantic content in accurate locations. For example, given an incomplete picture of a dog, we can guess that there is a tail, but we cannot determine its exact location. In this work, we propose to incorporate location uncertainty into MIM by using stochastic positional embeddings (StoP). Specifically, we condition the model on stochastic masked token positions drawn from a Gaussian distribution. StoP reduces overfitting to location features and guides the model toward learning features that are more robust to location uncertainties. Quantitatively, StoP improves downstream MIM performance on a variety of downstream tasks, including $+1.7\%$ on ImageNet linear probing using ViT-B, and $+2.5\%$ for ViT-H using $1\%$ of the data., Comment: Code and models available in https://github.com/amirbar/StoP

Self-supervised learning (SSL) algorithms can produce useful image representations by learning to associate different parts of natural images with one another. However, when taken to the extreme, SSL models can unintendedly memorize specific parts in individual training samples rather than learning semantically meaningful associations. In this work, we perform a systematic study of the unintended memorization of image-specific information in SSL models -- which we refer to as d\'ej\`a vu memorization. Concretely, we show that given the trained model and a crop of a training image containing only the background (e.g., water, sky, grass), it is possible to infer the foreground object with high accuracy or even visually reconstruct it. Furthermore, we show that d\'ej\`a vu memorization is common to different SSL algorithms, is exacerbated by certain design choices, and cannot be detected by conventional techniques for evaluating representation quality. Our study of d\'ej\`a vu memorization reveals previously unknown privacy risks in SSL models, as well as suggests potential practical mitigation strategies. Code is available at https://github.com/facebookresearch/DejaVu.

Joint-embedding based learning (e.g., SimCLR, MoCo, DINO) and reconstruction-based learning (e.g., BEiT, SimMIM, MAE) are the two leading paradigms for self-supervised learning of vision transformers, but they differ substantially in their transfer performance. Here, we aim to explain these differences by analyzing the impact of these objectives on the structure and transferability of the learned representations. Our analysis reveals that reconstruction-based learning features are significantly dissimilar to joint-embedding based learning features and that models trained with similar objectives learn similar features even across architectures. These differences arise early in the network and are primarily driven by attention and normalization layers. We find that joint-embedding features yield better linear probe transfer for classification because the different objectives drive different distributions of information and invariances in the learned representation. These differences explain opposite trends in transfer performance for downstream tasks that require spatial specificity in features. Finally, we address how fine-tuning changes reconstructive representations to enable better transfer, showing that fine-tuning re-organizes the information to be more similar to pre-trained joint embedding models.

Self-supervised learning, dubbed the dark matter of intelligence, is a promising path to advance machine learning. Yet, much like cooking, training SSL methods is a delicate art with a high barrier to entry. While many components are familiar, successfully training a SSL method involves a dizzying set of choices from the pretext tasks to training hyper-parameters. Our goal is to lower the barrier to entry into SSL research by laying the foundations and latest SSL recipes in the style of a cookbook. We hope to empower the curious researcher to navigate the terrain of methods, understand the role of the various knobs, and gain the know-how required to explore how delicious SSL can be.

Self-Supervised Learning (SSL) models rely on a pretext task to learn representations. Because this pretext task differs from the downstream tasks used to evaluate the performance of these models, there is an inherent misalignment or pretraining bias. A commonly used trick in SSL, shown to make deep networks more robust to such bias, is the addition of a small projector (usually a 2 or 3 layer multi-layer perceptron) on top of a backbone network during training. In contrast to previous work that studied the impact of the projector architecture, we here focus on a simpler, yet overlooked lever to control the information in the backbone representation. We show that merely changing its dimensionality -- by changing only the size of the backbone's very last block -- is a remarkably effective technique to mitigate the pretraining bias. It significantly improves downstream transfer performance for both Self-Supervised and Supervised pretrained models.

Joint Embedding Self-Supervised Learning (JE-SSL) has seen rapid developments in recent years, due to its promise to effectively leverage large unlabeled data. The development of JE-SSL methods was driven primarily by the search for ever increasing downstream classification accuracies, using huge computational resources, and typically built upon insights and intuitions inherited from a close parent JE-SSL method. This has led unwittingly to numerous pre-conceived ideas that carried over across methods e.g. that SimCLR requires very large mini batches to yield competitive accuracies; that strong and computationally slow data augmentations are required. In this work, we debunk several such ill-formed a priori ideas in the hope to unleash the full potential of JE-SSL free of unnecessary limitations. In fact, when carefully evaluating performances across different downstream tasks and properly optimizing hyper-parameters of the methods, we most often -- if not always -- see that these widespread misconceptions do not hold. For example we show that it is possible to train SimCLR to learn useful representations, while using a single image patch as negative example, and simple Gaussian noise as the only data augmentation for the positive pair. Along these lines, in the hope to democratize JE-SSL and to allow researchers to easily make more extensive evaluations of their methods, we introduce an optimized PyTorch library for SSL.

The framed standard model (FSM), constructed to explain the empirical mass and mixing patterns of quarks and leptons, gives the same result as the standard model in almost all areas in particle physics where it has been successfully applied, except for a few deviations such as the W mass and the g-2 of muons, where experiment is showing departures from what SM predicts. It predicts further the existence of a hidden sector of particles which may function as dark matter. In this paper, we first note that the above results involve the FSM undergoing a vacuum transition at a scale of around 17 MeV, where the vev's of the colour framons which are all nonzero above that scale acquire some vanishing components below it. This implies that the metric pertaining to these vanishing components would vanish also. Consequences should then ensue, mostly in the unknown hidden sector where empirical confirmation is hard at present to come by, but they give small reflections in the standard sector, some of which may have already been seen. However, one notes that if one imagines colour to be embedded, Kaluza-Klein fashion, into a higher dimensional space-time, then this VTR1 would cause 2 of the compactified dimensions to collapse. This might mean that when the universe cooled to the corresponding temperature of $10^{11}$K when it was about $10^{-3}$s old, this VTR1 collapse would cause the 3 spatial dimensions of the universe to expand to compensate. The resultant expansion is estimated, using FSM parameters previously determined from particle physics, to be capable, when extrapolated backwards in time, of bringing the present universe back inside the then horizon, solving thus formally the horizon problem. Besides, VTR1 being a global phenomenon in the FSM, it would switch on and off automatically and simultaneously over all space, thus requiring no additional strategy for a graceful exit., Comment: Inserted paragraph after eq. (4) and before equation (7), explaining the notation. Changed paragraph E1 in page 10. Added paragraph before the last in section 3. New references added 30 and 31

A successful paradigm in representation learning is to perform self-supervised pretraining using tasks based on mini-batch statistics (e.g., SimCLR, VICReg, SwAV, MSN). We show that in the formulation of all these methods is an overlooked prior to learn features that enable uniform clustering of the data. While this prior has led to remarkably semantic representations when pretraining on class-balanced data, such as ImageNet, we demonstrate that it can hamper performance when pretraining on class-imbalanced data. By moving away from conventional uniformity priors and instead preferring power-law distributed feature clusters, we show that one can improve the quality of the learned representations on real-world class-imbalanced datasets. To demonstrate this, we develop an extension of the Masked Siamese Networks (MSN) method to support the use of arbitrary features priors.

The Sample Adaptive Offset (SAO) filter has been introduced in HEVC to reduce general coding and banding artefacts in the reconstructed pictures, in complement to the De-Blocking Filter (DBF) which reduces artifacts at block boundaries specifically. The new video compression standard Versatile Video Coding (VVC) reduces the BD-rate by about 36% at the same reconstruction quality compared to HEVC. It implements an additional new in-loop Adaptive Loop Filter (ALF) on top of the DBF and the SAO filter, the latter remaining unchanged compared to HEVC. However, the relative performance of SAO in VVC has been lowered significantly. In this paper, it is proposed to revisit the SAO filter using Neural Networks (NN). The general principles of the SAO are kept, but the a-priori classification of SAO is replaced with a set of neural networks that determine which reconstructed samples should be corrected and in which proportion. Similarly to the original SAO, some parameters are determined at the encoder side and encoded per CTU. The average BD-rate gain of the proposed SAO improves VVC by at least 2.3% in Random Access while the overall complexity is kept relatively small compared to other NN-based methods.

One unexpected technique that emerged in recent years consists in training a Deep Network (DN) with a Self-Supervised Learning (SSL) method, and using this network on downstream tasks but with its last few projector layers entirely removed. This trick of throwing away the projector is actually critical for SSL methods to display competitive performances on ImageNet for which more than 30 percentage points can be gained that way. This is a little vexing, as one would hope that the network layer at which invariance is explicitly enforced by the SSL criterion during training (the last projector layer) should be the one to use for best generalization performance downstream. But it seems not to be, and this study sheds some light on why. This trick, which we name Guillotine Regularization (GR), is in fact a generically applicable method that has been used to improve generalization performance in transfer learning scenarios. In this work, we identify the underlying reasons behind its success and show that the optimal layer to use might change significantly depending on the training setup, the data or the downstream task. Lastly, we give some insights on how to reduce the need for a projector in SSL by aligning the pretext SSL task and the downstream task., Comment: Accepted at TMLR 2023

We show that, after resolving what was thought to be an ambiguity in the Higgs coupling, the FSM gives, apart from two extra terms (i) and (ii) to be specified below, an effective action in the standard sector which has the same form as the SM action, the two differing only in the values of the mass and mixing parameters of quarks and leptons which the SM takes as inputs from experiment while the FSM obtains as a result of a fit with a few parameters. Hence, to the accuracy that these two sets of parameters agree in value, and they do to a good extent as shown in earlier work, the FSM should give the same result as the SM in all the circumstances wherethe latter has been successfully applied, except for the noted modifications due to (i) and (ii). If so, it would be a big step forward for the FSM. The correction terms are: (i) a mixing between the SM's $\gamma - Z$ with a new vector boson in the hidden sector, (ii) a mixing between the standard Higgs with a new scalar boson also in the hidden sector. And these have been shown a few years back to lead to (i$'$) an enhancement of the $W$ mass over the SM value \cite{zmixed}, and (ii$'$) effects consistent with the $g - 2$ and some other anomalies, precisely the two deviations from the SM reported by experiments \cite{Wmassnew,g-2new} recently much in the news., Comment: Rectified a historical fact: What was said in version 1 to be a wrong deduction from FSM in an old paper (reference [10] of this version) was in fact not that, but a result from a older and looser framework R2M2 plus a further assumption. Title changed

We propose Masked Siamese Networks (MSN), a self-supervised learning framework for learning image representations. Our approach matches the representation of an image view containing randomly masked patches to the representation of the original unmasked image. This self-supervised pre-training strategy is particularly scalable when applied to Vision Transformers since only the unmasked patches are processed by the network. As a result, MSNs improve the scalability of joint-embedding architectures, while producing representations of a high semantic level that perform competitively on low-shot image classification. For instance, on ImageNet-1K, with only 5,000 annotated images, our base MSN model achieves 72.4% top-1 accuracy, and with 1% of ImageNet-1K labels, we achieve 75.7% top-1 accuracy, setting a new state-of-the-art for self-supervised learning on this benchmark. Our code is publicly available.

This paper presents a learning-based method to improve bi-prediction in video coding. In conventional video coding solutions, the motion compensation of blocks from already decoded reference pictures stands out as the principal tool used to predict the current frame. Especially, the bi-prediction, in which a block is obtained by averaging two different motion-compensated prediction blocks, significantly improves the final temporal prediction accuracy. In this context, we introduce a simple neural network that further improves the blending operation. A complexity balance, both in terms of network size and encoder mode selection, is carried out. Extensive tests on top of the recently standardized VVC codec are performed and show a BD-rate improvement of -1.4% in random access configuration for a network size of fewer than 10k parameters. We also propose a simple CPU-based implementation and direct network quantization to assess the complexity/gains tradeoff in a conventional codec framework.

Discovering what is learned by neural networks remains a challenge. In self-supervised learning, classification is the most common task used to evaluate how good a representation is. However, relying only on such downstream task can limit our understanding of what information is retained in the representation of a given input. In this work, we showcase the use of a Representation Conditional Diffusion Model (RCDM) to visualize in data space the representations learned by self-supervised models. The use of RCDM is motivated by its ability to generate high-quality samples -- on par with state-of-the-art generative models -- while ensuring that the representations of those samples are faithful i.e. close to the one used for conditioning. By using RCDM to analyze self-supervised models, we are able to clearly show visually that i) SSL (backbone) representation are not invariant to the data augmentations they were trained with -- thus debunking an often restated but mistaken belief; ii) SSL post-projector embeddings appear indeed invariant to these data augmentation, along with many other data symmetries; iii) SSL representations appear more robust to small adversarial perturbation of their inputs than representations trained in a supervised manner; and iv) that SSL-trained representations exhibit an inherent structure that can be explored thanks to RCDM visualization and enables image manipulation., Comment: Accepted at TMLR 2022