Randall Balestriero

Randall Balestriero, Mark Ibrahim, Vlad Sobal et al. · meta-ai

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.

Vishwanath Saragadam, Randall Balestriero, Ashok Veeraraghavan et al. · cmu

DeepTensor is a computationally efficient framework for low-rank decomposition of matrices and tensors using deep generative networks. We decompose a tensor as the product of low-rank tensor factors (e.g., a matrix as the outer product of two vectors), where each low-rank tensor is generated by a deep network (DN) that is trained in a self-supervised manner to minimize the mean-squared approximation error. Our key observation is that the implicit regularization inherent in DNs enables them to capture nonlinear signal structures (e.g., manifolds) that are out of the reach of classical linear methods like the singular value decomposition (SVD) and principal component analysis (PCA). Furthermore, in contrast to the SVD and PCA, whose performance deteriorates when the tensor's entries deviate from additive white Gaussian noise, we demonstrate that the performance of DeepTensor is robust to a wide range of distributions. We validate that DeepTensor is a robust and computationally efficient drop-in replacement for the SVD, PCA, nonnegative matrix factorization (NMF), and similar decompositions by exploring a range of real-world applications, including hyperspectral image denoising, 3D MRI tomography, and image classification. In particular, DeepTensor offers a 6dB signal-to-noise ratio improvement over standard denoising methods for signals corrupted by Poisson noise and learns to decompose 3D tensors 60 times faster than a single DN equipped with 3D convolutions.

Ravid Shwartz-Ziv, Randall Balestriero, Kenji Kawaguchi et al. · mit

Variance-Invariance-Covariance Regularization (VICReg) is a self-supervised learning (SSL) method that has shown promising results on a variety of tasks. However, the fundamental mechanisms underlying VICReg remain unexplored. In this paper, we present an information-theoretic perspective on the VICReg objective. We begin by deriving information-theoretic quantities for deterministic networks as an alternative to unrealistic stochastic network assumptions. We then relate the optimization of the VICReg objective to mutual information optimization, highlighting underlying assumptions and facilitating a constructive comparison with other SSL algorithms and derive a generalization bound for VICReg, revealing its inherent advantages for downstream tasks. Building on these results, we introduce a family of SSL methods derived from information-theoretic principles that outperform existing SSL techniques.

Randall Balestriero, Yann LeCun

Deep Neural Networks (DNNs) outshine alternative function approximators in many settings thanks to their modularity in composing any desired differentiable operator. The formed parametrized functional is then tuned to solve a task at hand from simple gradient descent. This modularity comes at the cost of making strict enforcement of constraints on DNNs, e.g. from a priori knowledge of the task, or from desired physical properties, an open challenge. In this paper we propose the first provable affine constraint enforcement method for DNNs that only requires minimal changes into a given DNN's forward-pass, that is computationally friendly, and that leaves the optimization of the DNN's parameter to be unconstrained, i.e. standard gradient-based method can be employed. Our method does not require any sampling and provably ensures that the DNN fulfills the affine constraint on a given input space's region at any point during training, and testing. We coin this method POLICE, standing for Provably Optimal LInear Constraint Enforcement. Github: https://github.com/RandallBalestriero/POLICE

Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero et al. · meta-ai

Deep learning vision systems are widely deployed across applications where reliability is critical. However, even today's best models can fail to recognize an object when its pose, lighting, or background varies. While existing benchmarks surface examples challenging for models, they do not explain why such mistakes arise. To address this need, we introduce ImageNet-X, a set of sixteen human annotations of factors such as pose, background, or lighting the entire ImageNet-1k validation set as well as a random subset of 12k training images. Equipped with ImageNet-X, we investigate 2,200 current recognition models and study the types of mistakes as a function of model's (1) architecture, e.g. transformer vs. convolutional, (2) learning paradigm, e.g. supervised vs. self-supervised, and (3) training procedures, e.g., data augmentation. Regardless of these choices, we find models have consistent failure modes across ImageNet-X categories. We also find that while data augmentation can improve robustness to certain factors, they induce spill-over effects to other factors. For example, strong random cropping hurts robustness on smaller objects. Together, these insights suggest to advance the robustness of modern vision models, future research should focus on collecting additional data and understanding data augmentation schemes. Along with these insights, we release a toolkit based on ImageNet-X to spur further study into the mistakes image recognition systems make.

Randall Balestriero, Yann LeCun

Self-Supervised Learning (SSL) surmises that inputs and pairwise positive relationships are enough to learn meaningful representations. Although SSL has recently reached a milestone: outperforming supervised methods in many modalities\dots the theoretical foundations are limited, method-specific, and fail to provide principled design guidelines to practitioners. In this paper, we propose a unifying framework under the helm of spectral manifold learning to address those limitations. Through the course of this study, we will rigorously demonstrate that VICReg, SimCLR, BarlowTwins et al. correspond to eponymous spectral methods such as Laplacian Eigenmaps, Multidimensional Scaling et al. This unification will then allow us to obtain (i) the closed-form optimal representation for each method, (ii) the closed-form optimal network parameters in the linear regime for each method, (iii) the impact of the pairwise relations used during training on each of those quantities and on downstream task performances, and most importantly, (iv) the first theoretical bridge between contrastive and non-contrastive methods towards global and local spectral embedding methods respectively, hinting at the benefits and limitations of each. For example, (i) if the pairwise relation is aligned with the downstream task, any SSL method can be employed successfully and will recover the supervised method, but in the low data regime, VICReg's invariance hyper-parameter should be high; (ii) if the pairwise relation is misaligned with the downstream task, VICReg with small invariance hyper-parameter should be preferred over SimCLR or BarlowTwins.

Quentin Garrido, Randall Balestriero, Laurent Najman et al.

Joint-Embedding Self Supervised Learning (JE-SSL) has seen a rapid development, with the emergence of many method variations but only few principled guidelines that would help practitioners to successfully deploy them. The main reason for that pitfall comes from JE-SSL's core principle of not employing any input reconstruction therefore lacking visual cues of unsuccessful training. Adding non informative loss values to that, it becomes difficult to deploy SSL on a new dataset for which no labels can help to judge the quality of the learned representation. In this study, we develop a simple unsupervised criterion that is indicative of the quality of the learned JE-SSL representations: their effective rank. Albeit simple and computationally friendly, this method -- coined RankMe -- allows one to assess the performance of JE-SSL representations, even on different downstream datasets, without requiring any labels. A further benefit of RankMe is that it does not have any training or hyper-parameters to tune. Through thorough empirical experiments involving hundreds of training episodes, we demonstrate how RankMe can be used for hyperparameter selection with nearly no reduction in final performance compared to the current selection method that involve a dataset's labels. We hope that RankMe will facilitate the deployment of JE-SSL towards domains that do not have the opportunity to rely on labels for representations' quality assessment.

Randall Balestriero, Leon Bottou, Yann LeCun

Regularization is a fundamental technique to prevent over-fitting and to improve generalization performances by constraining a model's complexity. Current Deep Networks heavily rely on regularizers such as Data-Augmentation (DA) or weight-decay, and employ structural risk minimization, i.e. cross-validation, to select the optimal regularization hyper-parameters. In this study, we demonstrate that techniques such as DA or weight decay produce a model with a reduced complexity that is unfair across classes. The optimal amount of DA or weight decay found from cross-validation leads to disastrous model performances on some classes e.g. on Imagenet with a resnet50, the "barn spider" classification test accuracy falls from $68\%$ to $46\%$ only by introducing random crop DA during training. Even more surprising, such performance drop also appears when introducing uninformative regularization techniques such as weight decay. Those results demonstrate that our search for ever increasing generalization performance -- averaged over all classes and samples -- has left us with models and regularizers that silently sacrifice performances on some classes. This scenario can become dangerous when deploying a model on downstream tasks e.g. an Imagenet pre-trained resnet50 deployed on INaturalist sees its performances fall from $70\%$ to $30\%$ on class \#8889 when introducing random crop DA during the Imagenet pre-training phase. Those results demonstrate that designing novel regularizers without class-dependent bias remains an open research question.

Vivien Cabannes, Bobak T. Kiani, Randall Balestriero et al.

Self-supervised learning (SSL) has emerged as a powerful framework to learn representations from raw data without supervision. Yet in practice, engineers face issues such as instability in tuning optimizers and collapse of representations during training. Such challenges motivate the need for a theory to shed light on the complex interplay between the choice of data augmentation, network architecture, and training algorithm. We study such an interplay with a precise analysis of generalization performance on both pretraining and downstream tasks in a theory friendly setup, and highlight several insights for SSL practitioners that arise from our theory.

Bobak Kiani, Randall Balestriero, Yann LeCun et al.

In learning with recurrent or very deep feed-forward networks, employing unitary matrices in each layer can be very effective at maintaining long-range stability. However, restricting network parameters to be unitary typically comes at the cost of expensive parameterizations or increased training runtime. We propose instead an efficient method based on rank-$k$ updates -- or their rank-$k$ approximation -- that maintains performance at a nearly optimal training runtime. We introduce two variants of this method, named Direct (projUNN-D) and Tangent (projUNN-T) projected Unitary Neural Networks, that can parameterize full $N$-dimensional unitary or orthogonal matrices with a training runtime scaling as $O(kN^2)$. Our method either projects low-rank gradients onto the closest unitary matrix (projUNN-T) or transports unitary matrices in the direction of the low-rank gradient (projUNN-D). Even in the fastest setting ($k=1$), projUNN is able to train a model's unitary parameters to reach comparable performances against baseline implementations. In recurrent neural network settings, projUNN closely matches or exceeds benchmarked results from prior unitary neural networks. Finally, we preliminarily explore projUNN in training orthogonal convolutional neural networks, which are currently unable to outperform state of the art models but can potentially enhance stability and robustness at large depth.

Ahmed Imtiaz Humayun, Randall Balestriero, Richard Baraniuk

We present Polarity Sampling, a theoretically justified plug-and-play method for controlling the generation quality and diversity of pre-trained deep generative networks DGNs). Leveraging the fact that DGNs are, or can be approximated by, continuous piecewise affine splines, we derive the analytical DGN output space distribution as a function of the product of the DGN's Jacobian singular values raised to a power $ρ$. We dub $ρ$ the $\textbf{polarity}$ parameter and prove that $ρ$ focuses the DGN sampling on the modes ($ρ< 0$) or anti-modes ($ρ> 0$) of the DGN output-space distribution. We demonstrate that nonzero polarity values achieve a better precision-recall (quality-diversity) Pareto frontier than standard methods, such as truncation, for a number of state-of-the-art DGNs. We also present quantitative and qualitative results on the improvement of overall generation quality (e.g., in terms of the Frechet Inception Distance) for a number of state-of-the-art DGNs, including StyleGAN3, BigGAN-deep, NVAE, for different conditional and unconditional image generation tasks. In particular, Polarity Sampling redefines the state-of-the-art for StyleGAN2 on the FFHQ Dataset to FID 2.57, StyleGAN2 on the LSUN Car Dataset to FID 2.27 and StyleGAN3 on the AFHQv2 Dataset to FID 3.95. Demo: bit.ly/polarity-samp

Ahmed Imtiaz Humayun, Randall Balestriero, Guha Balakrishnan et al.

Current Deep Network (DN) visualization and interpretability methods rely heavily on data space visualizations such as scoring which dimensions of the data are responsible for their associated prediction or generating new data features or samples that best match a given DN unit or representation. In this paper, we go one step further by developing the first provably exact method for computing the geometry of a DN's mapping - including its decision boundary - over a specified region of the data space. By leveraging the theory of Continuous Piece-Wise Linear (CPWL) spline DNs, SplineCam exactly computes a DNs geometry without resorting to approximations such as sampling or architecture simplification. SplineCam applies to any DN architecture based on CPWL nonlinearities, including (leaky-)ReLU, absolute value, maxout, and max-pooling and can also be applied to regression DNs such as implicit neural representations. Beyond decision boundary visualization and characterization, SplineCam enables one to compare architectures, measure generalizability and sample from the decision boundary on or off the manifold. Project Website: bit.ly/splinecam.

Ravid Shwartz-Ziv, Randall Balestriero, Yann LeCun

In this paper, we examine self-supervised learning methods, particularly VICReg, to provide an information-theoretical understanding of their construction. As a first step, we demonstrate how information-theoretic quantities can be obtained for a deterministic network, offering a possible alternative to prior work that relies on stochastic models. This enables us to demonstrate how VICReg can be (re)discovered from first principles and its assumptions about data distribution. Furthermore, we empirically demonstrate the validity of our assumptions, confirming our novel understanding of VICReg. Finally, we believe that the derivation and insights we obtain can be generalized to many other SSL methods, opening new avenues for theoretical and practical understanding of SSL and transfer learning.

Randall Balestriero, Richard G. Baraniuk

A critically important, ubiquitous, and yet poorly understood ingredient in modern deep networks (DNs) is batch normalization (BN), which centers and normalizes the feature maps. To date, only limited progress has been made understanding why BN boosts DN learning and inference performance; work has focused exclusively on showing that BN smooths a DN's loss landscape. In this paper, we study BN theoretically from the perspective of function approximation; we exploit the fact that most of today's state-of-the-art DNs are continuous piecewise affine (CPA) splines that fit a predictor to the training data via affine mappings defined over a partition of the input space (the so-called "linear regions"). {\em We demonstrate that BN is an unsupervised learning technique that -- independent of the DN's weights or gradient-based learning -- adapts the geometry of a DN's spline partition to match the data.} BN provides a "smart initialization" that boosts the performance of DN learning, because it adapts even a DN initialized with random weights to align its spline partition with the data. We also show that the variation of BN statistics between mini-batches introduces a dropout-like random perturbation to the partition boundaries and hence the decision boundary for classification problems. This per mini-batch perturbation reduces overfitting and improves generalization by increasing the margin between the training samples and the decision boundary.

Bobak T. Kiani, Randall Balestriero, Yubei Chen et al.

The fundamental goal of self-supervised learning (SSL) is to produce useful representations of data without access to any labels for classifying the data. Modern methods in SSL, which form representations based on known or constructed relationships between samples, have been particularly effective at this task. Here, we aim to extend this framework to incorporate algorithms based on kernel methods where embeddings are constructed by linear maps acting on the feature space of a kernel. In this kernel regime, we derive methods to find the optimal form of the output representations for contrastive and non-contrastive loss functions. This procedure produces a new representation space with an inner product denoted as the induced kernel which generally correlates points which are related by an augmentation in kernel space and de-correlates points otherwise. We analyze our kernel model on small datasets to identify common features of self-supervised learning algorithms and gain theoretical insights into their performance on downstream tasks.

Vivien Cabannes, Leon Bottou, Yann Lecun et al.

Self-Supervised Learning (SSL) has emerged as the solution of choice to learn transferable representations from unlabeled data. However, SSL requires to build samples that are known to be semantically akin, i.e. positive views. Requiring such knowledge is the main limitation of SSL and is often tackled by ad-hoc strategies e.g. applying known data-augmentations to the same input. In this work, we formalize and generalize this principle through Positive Active Learning (PAL) where an oracle queries semantic relationships between samples. PAL achieves three main objectives. First, it unveils a theoretically grounded learning framework beyond SSL, based on similarity graphs, that can be extended to tackle supervised and semi-supervised learning depending on the employed oracle. Second, it provides a consistent algorithm to embed a priori knowledge, e.g. some observed labels, into any SSL losses without any change in the training pipeline. Third, it provides a proper active learning framework yielding low-cost solutions to annotate datasets, arguably bringing the gap between theory and practice of active learning that is based on simple-to-answer-by-non-experts queries of semantic relationships between inputs.

Grégoire Mialon, Randall Balestriero, Yann LeCun

Self-Supervised Learning (SSL) methods such as VICReg, Barlow Twins or W-MSE avoid collapse of their joint embedding architectures by constraining or regularizing the covariance matrix of their projector's output. This study highlights important properties of such strategy, which we coin Variance-Covariance regularization (VCReg). More precisely, we show that {\em VCReg combined to a MLP projector enforces pairwise independence between the features of the learned representation}. This result emerges by bridging VCReg applied on the projector's output to kernel independence criteria applied on the projector's input. We empirically validate our findings where (i) we put in evidence which projector's characteristics favor pairwise independence, (ii) we demonstrate pairwise independence to be beneficial for out-of-domain generalization, (iii) we demonstrate that the scope of VCReg goes beyond SSL by using it to solve Independent Component Analysis. This provides the first theoretical motivation and explanation of MLP projectors in SSL.

Rudolf H. Riedi, Randall Balestriero, Richard G. Baraniuk

We develop new theoretical results on matrix perturbation to shed light on the impact of architecture on the performance of a deep network. In particular, we explain analytically what deep learning practitioners have long observed empirically: the parameters of some deep architectures (e.g., residual networks, ResNets, and Dense networks, DenseNets) are easier to optimize than others (e.g., convolutional networks, ConvNets). Building on our earlier work connecting deep networks with continuous piecewise-affine splines, we develop an exact local linear representation of a deep network layer for a family of modern deep networks that includes ConvNets at one end of a spectrum and ResNets, DenseNets, and other networks with skip connections at the other. For regression and classification tasks that optimize the squared-error loss, we show that the optimization loss surface of a modern deep network is piecewise quadratic in the parameters, with local shape governed by the singular values of a matrix that is a function of the local linear representation. We develop new perturbation results for how the singular values of matrices of this sort behave as we add a fraction of the identity and multiply by certain diagonal matrices. A direct application of our perturbation results explains analytically why a network with skip connections (such as a ResNet or DenseNet) is easier to optimize than a ConvNet: thanks to its more stable singular values and smaller condition number, the local loss surface of such a network is less erratic, less eccentric, and features local minima that are more accommodating to gradient-based optimization. Our results also shed new light on the impact of different nonlinear activation functions on a deep network's singular values, regardless of its architecture.

Mahmoud Assran, Randall Balestriero, Quentin Duval et al.

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.

Florian Bordes, Randall Balestriero, Quentin Garrido et al.

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.

Florian Bordes, Randall Balestriero, Pascal Vincent

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.

Randall Balestriero, Yann LeCun

Learning manipulable representations of the world and its dynamics is central to AI. Joint-Embedding Predictive Architectures (JEPAs) offer a promising blueprint, but lack of practical guidance and theory has led to ad-hoc R&D. We present a comprehensive theory of JEPAs and instantiate it in {\bf LeJEPA}, a lean, scalable, and theoretically grounded training objective. First, we identify the isotropic Gaussian as the optimal distribution that JEPAs' embeddings should follow to minimize downstream prediction risk. Second, we introduce a novel objective--{\bf Sketched Isotropic Gaussian Regularization} (SIGReg)--to constrain embeddings to reach that ideal distribution. Combining the JEPA predictive loss with SIGReg yields LeJEPA with numerous theoretical and practical benefits: (i) single trade-off hyperparameter, (ii) linear time and memory complexity, (iii) stability across hyper-parameters, architectures (ResNets, ViTs, ConvNets) and domains, (iv) heuristics-free, e.g., no stop-gradient, no teacher-student, no hyper-parameter schedulers, and (v) distributed training-friendly implementation requiring only $\approx$50 lines of code. Our empirical validation covers 10+ datasets, 60+ architectures, all with varying scales and domains. As an example, using imagenet-1k for pretraining and linear evaluation with frozen backbone, LeJEPA reaches 79\% with a ViT-H/14. We hope that the simplicity and theory-friendly ecosystem offered by LeJEPA will reestablish self-supervised pre-training as a core pillar of AI research (\href{https://github.com/rbalestr-lab/lejepa}{GitHub repo}).

Ahmed Imtiaz Humayun, Randall Balestriero, Anastasios Kyrillidis et al.

Centroid based clustering methods such as k-means, k-medoids and k-centers are heavily applied as a go-to tool in exploratory data analysis. In many cases, those methods are used to obtain representative centroids of the data manifold for visualization or summarization of a dataset. Real world datasets often contain inherent abnormalities, e.g., repeated samples and sampling bias, that manifest imbalanced clustering. We propose to remedy such a scenario by introducing a maximal radius constraint $r$ on the clusters formed by the centroids, i.e., samples from the same cluster should not be more than $2r$ apart in terms of $\ell_2$ distance. We achieve this constraint by solving a semi-definite program, followed by a linear assignment problem with quadratic constraints. Through qualitative results, we show that our proposed method is robust towards dataset imbalances and sampling artifacts. To the best of our knowledge, ours is the first constrained k-means clustering method with hard radius constraints. Codes at https://bit.ly/kmeans-constrained

Florian Bordes, Samuel Lavoie, Randall Balestriero et al.

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.

Haider Al-Tahan, Quentin Garrido, Randall Balestriero et al.

Significant research efforts have been made to scale and improve vision-language model (VLM) training approaches. Yet, with an ever-growing number of benchmarks, researchers are tasked with the heavy burden of implementing each protocol, bearing a non-trivial computational cost, and making sense of how all these benchmarks translate into meaningful axes of progress. To facilitate a systematic evaluation of VLM progress, we introduce UniBench: a unified implementation of 50+ VLM benchmarks spanning a comprehensive range of carefully categorized capabilities from object recognition to spatial awareness, counting, and much more. We showcase the utility of UniBench for measuring progress by evaluating nearly 60 publicly available vision-language models, trained on scales of up to 12.8B samples. We find that while scaling training data or model size can boost many vision-language model capabilities, scaling offers little benefit for reasoning or relations. Surprisingly, we also discover today's best VLMs struggle on simple digit recognition and counting tasks, e.g. MNIST, which much simpler networks can solve. Where scale falls short, we find that more precise interventions, such as data quality or tailored-learning objectives offer more promise. For practitioners, we also offer guidance on selecting a suitable VLM for a given application. Finally, we release an easy-to-run UniBench code-base with the full set of 50+ benchmarks and comparisons across 59 models as well as a distilled, representative set of benchmarks that runs in 5 minutes on a single GPU.

Wei-Yin Ko, Daniel D'souza, Karina Nguyen et al.

Ensembling multiple Deep Neural Networks (DNNs) is a simple and effective way to improve top-line metrics and to outperform a larger single model. In this work, we go beyond top-line metrics and instead explore the impact of ensembling on subgroup performances. Surprisingly, we observe that even with a simple homogeneous ensemble -- all the individual DNNs share the same training set, architecture, and design choices -- the minority group performance disproportionately improves with the number of models compared to the majority group, i.e. fairness naturally emerges from ensembling. Even more surprising, we find that this gain keeps occurring even when a large number of models is considered, e.g. $20$, despite the fact that the average performance of the ensemble plateaus with fewer models. Our work establishes that simple DNN ensembles can be a powerful tool for alleviating disparate impact from DNN classifiers, thus curbing algorithmic harm. We also explore why this is the case. We find that even in homogeneous ensembles, varying the sources of stochasticity through parameter initialization, mini-batch sampling, and data-augmentation realizations, results in different fairness outcomes.

Vlad Sobal, Mark Ibrahim, Randall Balestriero et al.

Learning good representations involves capturing the diverse ways in which data samples relate. Contrastive loss - an objective matching related samples - underlies methods from self-supervised to multimodal learning. Contrastive losses, however, can be viewed more broadly as modifying a similarity graph to indicate how samples should relate in the embedding space. This view reveals a shortcoming in contrastive learning: the similarity graph is binary, as only one sample is the related positive sample. Crucially, similarities \textit{across} samples are ignored. Based on this observation, we revise the standard contrastive loss to explicitly encode how a sample relates to others. We experiment with this new objective, called $\mathbb{X}$-Sample Contrastive, to train vision models based on similarities in class or text caption descriptions. Our study spans three scales: ImageNet-1k with 1 million, CC3M with 3 million, and CC12M with 12 million samples. The representations learned via our objective outperform both contrastive self-supervised and vision-language models trained on the same data across a range of tasks. When training on CC12M, we outperform CLIP by $0.6\%$ on both ImageNet and ImageNet Real. Our objective appears to work particularly well in lower-data regimes, with gains over CLIP of $16.8\%$ on ImageNet and $18.1\%$ on ImageNet Real when training with CC3M. Finally, our objective seems to encourage the model to learn representations that separate objects from their attributes and backgrounds, with gains of $3.3$-$5.6$\% over CLIP on ImageNet9. We hope the proposed solution takes a small step towards developing richer learning objectives for understanding sample relations in foundation models.

Vivien Cabannes, Alberto Bietti, Randall Balestriero

Unsupervised representation learning aims at describing raw data efficiently to solve various downstream tasks. It has been approached with many techniques, such as manifold learning, diffusion maps, or more recently self-supervised learning. Those techniques are arguably all based on the underlying assumption that target functions, associated with future downstream tasks, have low variations in densely populated regions of the input space. Unveiling minimal variations as a guiding principle behind unsupervised representation learning paves the way to better practical guidelines for self-supervised learning algorithms.

Ahmed Imtiaz Humayun, Randall Balestriero, Richard Baraniuk

The study of Deep Network (DN) training dynamics has largely focused on the evolution of the loss function, evaluated on or around train and test set data points. In fact, many DN phenomenon were first introduced in literature with that respect, e.g., double descent, grokking. In this study, we look at the training dynamics of the input space partition or linear regions formed by continuous piecewise affine DNs, e.g., networks with (leaky)ReLU nonlinearities. First, we present a novel statistic that encompasses the local complexity (LC) of the DN based on the concentration of linear regions inside arbitrary dimensional neighborhoods around data points. We observe that during training, the LC around data points undergoes a number of phases, starting with a decreasing trend after initialization, followed by an ascent and ending with a final descending trend. Using exact visualization methods, we come across the perplexing observation that during the final LC descent phase of training, linear regions migrate away from training and test samples towards the decision boundary, making the DN input-output nearly linear everywhere else. We also observe that the different LC phases are closely related to the memorization and generalization performance of the DN, especially during grokking.

Xue Xia, Randall Balestriero, Tao Zhang et al.

Historical maps are unique and valuable archives that document geographic features across different time periods. However, automated analysis of historical map images remains a significant challenge due to their wide stylistic variability and the scarcity of annotated training data. Constructing linked spatio-temporal datasets from historical map time series is even more time-consuming and labor-intensive, as it requires synthesizing information from multiple maps. Such datasets are essential for applications such as dating buildings, analyzing the development of road networks and settlements, studying environmental changes etc. We present MapSAM2, a unified framework for automatically segmenting both historical map images and time series. Built on a visual foundation model, MapSAM2 adapts to diverse segmentation tasks with few-shot fine-tuning. Our key innovation is to treat both historical map images and time series as videos. For images, we process a set of tiles as a video, enabling the memory attention mechanism to incorporate contextual cues from similar tiles, leading to improved geometric accuracy, particularly for areal features. For time series, we introduce the annotated Siegfried Building Time Series Dataset and, to reduce annotation costs, propose generating pseudo time series from single-year maps by simulating common temporal transformations. Experimental results show that MapSAM2 learns temporal associations effectively and can accurately segment and link buildings in time series under limited supervision or using pseudo videos. We will release both our dataset and code to support future research.

Randall Balestriero, Romain Cosentino, Sarath Shekkizhar

Large Language Models (LLMs) drive current AI breakthroughs despite very little being known about their internal representations. In this work, we propose to shed the light on LLMs inner mechanisms through the lens of geometry. In particular, we develop in closed form $(i)$ the intrinsic dimension in which the Multi-Head Attention embeddings are constrained to exist and $(ii)$ the partition and per-region affine mappings of the feedforward (MLP) network of LLMs' layers. Our theoretical findings further enable the design of novel principled solutions applicable to state-of-the-art LLMs. First, we show that, through our geometric understanding, we can bypass LLMs' RLHF protection by controlling the embedding's intrinsic dimension through informed prompt manipulation. Second, we derive interpretable geometrical features that can be extracted from any (pre-trained) LLM, providing a rich abstract representation of their inputs. We observe that these features are sufficient to help solve toxicity detection, and even allow the identification of various types of toxicity. Our results demonstrate how, even in large-scale regimes, exact theoretical results can answer practical questions in LLMs. Code: https://github.com/RandallBalestriero/SplineLLM

Randall Balestriero

Costly, noisy, and over-specialized, labels are to be set aside in favor of unsupervised learning if we hope to learn cheap, reliable, and transferable models. To that end, spectral embedding, self-supervised learning, or generative modeling have offered competitive solutions. Those methods however come with numerous challenges \textit{e.g.} estimating geodesic distances, specifying projector architectures and anti-collapse losses, or specifying decoder architectures and reconstruction losses. In contrast, we introduce a simple explainable alternative -- coined \textbf{DIET} -- to learn representations from unlabeled data, free of those challenges. \textbf{DIET} is blatantly simple: take one's favorite classification setup and use the \textbf{D}atum \textbf{I}nd\textbf{E}x as its \textbf{T}arget class, \textit{i.e. each sample is its own class}, no further changes needed. \textbf{DIET} works without a decoder/projector network, is not based on positive pairs nor reconstruction, introduces no hyper-parameters, and works out-of-the-box across datasets and architectures. Despite \textbf{DIET}'s simplicity, the learned representations are of high-quality and often on-par with the state-of-the-art \textit{e.g.} using a linear classifier on top of DIET's learned representation reaches $71.4\%$ on CIFAR100 with a Resnet101, $52.5\%$ on TinyImagenet with a Resnext50.

Hai Huang, Yann LeCun, Randall Balestriero

Large Language Model (LLM) pretraining, finetuning, and evaluation rely on input-space reconstruction and generative capabilities. Yet, it has been observed in vision that embedding-space training objectives, e.g., with Joint Embedding Predictive Architectures (JEPAs), are far superior to their input-space counterpart. That mismatch in how training is achieved between language and vision opens up a natural question: {\em can language training methods learn a few tricks from the vision ones?} The lack of JEPA-style LLM is a testimony of the challenge in designing such objectives for language. In this work, we propose a first step in that direction where we develop LLM-JEPA, a JEPA based solution for LLMs applicable both to finetuning and pretraining. Thus far, LLM-JEPA is able to outperform the standard LLM training objectives by a significant margin across models, all while being robust to overfiting. Those findings are observed across numerous datasets (NL-RX, GSM8K, Spider, RottenTomatoes) and various models from the Llama3, OpenELM, Gemma2 and Olmo families. Code: https://github.com/rbalestr-lab/llm-jepa.

Thomas Walker, Ahmed Imtiaz Humayun, Randall Balestriero et al.

A key challenge for the machine learning community is to understand and accelerate the training dynamics of deep networks that lead to delayed generalisation and emergent robustness to input perturbations, also known as grokking. Prior work has associated phenomena like delayed generalisation with the transition of a deep network from a linear to a feature learning regime, and emergent robustness with changes to the network's functional geometry, in particular the arrangement of the so-called linear regions in deep networks employing continuous piecewise affine nonlinearities. Here, we explain how grokking is realised in the Jacobian of a deep network and demonstrate that aligning a network's Jacobians with the training data (in the sense of cosine similarity) ensures grokking under a low-rank Jacobian assumption. Our results provide a strong theoretical motivation for the use of Jacobian regularisation in optimizing deep networks -- a method we introduce as GrokAlign -- which we show empirically to induce grokking much sooner than more conventional regularizers like weight decay. Moreover, we introduce centroid alignment as a tractable and interpretable simplification of Jacobian alignment that effectively identifies and tracks the stages of deep network training dynamics. Accompanying webpage (https://thomaswalker1.github.io/blog/grokalign.html) and code (https://github.com/ThomasWalker1/grokalign).

Wenquan Lu, Jiaqi Zhang, Hugues Van Assel et al.

Self-Supervised Learning (SSL) has become a powerful solution to extract rich representations from unlabeled data. Yet, SSL research is mostly focused on clean, curated and high-quality datasets. As a result, applying SSL on noisy data remains a challenge, despite being crucial to applications such as astrophysics, medical imaging, geophysics or finance. In this work, we present a fully self-supervised framework that enables noise-robust representation learning without requiring a denoiser at inference or downstream fine-tuning. Our method first trains an SSL denoiser on noisy data, then uses it to construct a denoised-to-noisy data curriculum (i.e., training first on denoised, then noisy samples) for pretraining a SSL backbone (e.g., DINOv2), combined with a teacher-guided regularization that anchors noisy embeddings to their denoised counterparts. This process encourages the model to internalize noise robustness. Notably, the denoiser can be discarded after pretraining, simplifying deployment. On ImageNet-1k with ViT-B under extreme Gaussian noise ($σ=255$, SNR = 0.72 dB), our method improves linear probing accuracy by 4.8% over DINOv2, demonstrating that denoiser-free robustness can emerge from noise-aware pretraining. The code is available at https://github.com/wenquanlu/noisy_dinov2.

Karen Ullrich, Jingtong Su, Claudia Shi et al.

Reliability is key to realizing the promise of autonomous UI-Agents, multimodal agents that directly interact with apps in the same manner as humans, as users must be able to trust an agent to complete a given task. Current evaluations rely on fixed environments, often clones of existing apps, which are limited in that they can only shed light on whether or how often an agent can complete a task within a specific environment. When deployed however, agents are likely to encounter variations in app design and content that can affect an agent's ability to complete a task. To address this blind spot of measuring agent reliability across app variations, we develop OpenApps, a light-weight open-source ecosystem with six apps (messenger, calendar, maps, etc.) that are configurable in appearance and content. OpenApps requires just a single CPU to run, enabling easy generation and deployment of thousands of versions of each app. Specifically, we run more than 10,000 independent evaluations to study reliability across seven leading multimodal agents. We find that while standard reliability within a fixed app is relatively stable, reliability can vary drastically when measured across app variations. Task success rates for many agents can fluctuate by more than $50\%$ across app variations. For example, Kimi-VL-3B's average success across all tasks fluctuates from $63\%$ to just $4\%$ across app versions. We also find agent behaviors such as looping or hallucinating actions can differ drastically depending on the environment configuration. These initial findings highlight the importance of measuring reliability along this new dimension of app variations. OpenApps is available at https://facebookresearch.github.io/OpenApps/

Camilo Tamayo-Rousseau, Yunjia Zhao, Yiqun Zhang et al.

Self-attention mechanisms are foundational to Transformer architectures, supporting their impressive success in a wide range of tasks. While there are many self-attention variants, their robustness to noise and spurious correlations has not been well studied. This study evaluates Softmax, Sigmoid, Linear, Doubly Stochastic, and Cosine attention within Vision Transformers under different data corruption scenarios. Through testing across the CIFAR-10, CIFAR-100, and Imagenette datasets, we show that Doubly Stochastic attention is the most robust. It consistently outperformed the next best mechanism by $0.1\%-5.1\%$ when training data, or both training and testing data, were corrupted. Our findings inform self-attention selection in contexts with imperfect data. The code used is available at https://github.com/ctamayor/NeurIPS-Robustness-ViT.

Jiaqi Zhang, Juntuo Wang, Zhixin Sun et al.

Large-scale vision foundation models such as DINOv2 boast impressive performances by leveraging massive architectures and training datasets. But numerous scenarios require practitioners to reproduce those pre-training solutions, such as on private data, new modalities, or simply for scientific questioning--which is currently extremely demanding computation-wise. We thus propose a novel pre-training strategy for DINOv2 that simultaneously accelerates convergence--and strengthens robustness to common corruptions as a by-product. Our approach involves a frequency filtering curriculum--low-frequency being seen first--and the Gaussian noise patching augmentation. Applied to a ViT-B/16 backbone trained on ImageNet-1K, while pre-training time and FLOPs are reduced by 1.6x and 2.25x, our method still achieves matching robustness in corruption benchmarks (ImageNet-C) and maintains competitive linear probing performance compared with baseline. This dual benefit of efficiency and robustness makes large-scale self-supervised foundation modeling more attainable, while opening the door to novel exploration around data curriculum and augmentation as means to improve self-supervised learning models robustness. The code is available at https://github.com/KevinZ0217/fast_dinov2

Jeremy Budd, Javier Ideami, Benjamin Macdowall Rynne et al.

Sparse autoencoders (SAEs) have received considerable recent attention as tools for mechanistic interpretability, showing success at extracting interpretable features even from very large LLMs. However, this research has been largely empirical, and there have been recent doubts about the true utility of SAEs. In this work, we seek to enhance the theoretical understanding of SAEs, using the spline theory of deep learning. By situating SAEs in this framework: we discover that SAEs generalise ``$k$-means autoencoders'' to be piecewise affine, but sacrifice accuracy for interpretability vs. the optimal ``$k$-means-esque plus local principal component analysis (PCA)'' piecewise affine autoencoder. We characterise the underlying geometry of (TopK) SAEs using power diagrams. And we develop a novel proximal alternating method SGD (PAM-SGD) algorithm for training SAEs, with both solid theoretical foundations and promising empirical results in MNIST and LLM experiments, particularly in sample efficiency and (in the LLM setting) improved sparsity of codes. All code is available at: https://github.com/splInterp2025/splInterp

Omer Ronen, Ahmed Imtiaz Humayun, Richard Baraniuk et al.

We develop Latent Exploration Score (LES) to mitigate over-exploration in Latent Space Optimization (LSO), a popular method for solving black-box discrete optimization problems. LSO utilizes continuous optimization within the latent space of a Variational Autoencoder (VAE) and is known to be susceptible to over-exploration, which manifests in unrealistic solutions that reduce its practicality. LES leverages the trained decoder's approximation of the data distribution, and can be employed with any VAE decoder - including pretrained ones - without additional training, architectural changes or access to the training data. Our evaluation across five LSO benchmark tasks and twenty-two VAE models demonstrates that LES always enhances the quality of the solutions while maintaining high objective values, leading to improvements over existing solutions in most cases. We believe that new avenues to LSO will be opened by LES' ability to identify out of distribution areas, differentiability, and computational tractability. Open source code for LES is available at https://github.com/OmerRonen/les.

Nicholas Dronen, Randall Balestriero

Catastrophic forgetting -- the phenomenon of a neural network learning a task t1 and losing the ability to perform it after being trained on some other task t2 -- is a long-standing problem for neural networks [McCloskey and Cohen, 1989]. We present a method, Eidetic Learning, that provably solves catastrophic forgetting. A network trained with Eidetic Learning -- here, an EideticNet -- requires no rehearsal or replay. We consider successive discrete tasks and show how at inference time an EideticNet automatically routes new instances without auxiliary task information. An EideticNet bears a family resemblance to the sparsely-gated Mixture-of-Experts layer Shazeer et al. [2016] in that network capacity is partitioned across tasks and the network itself performs data-conditional routing. An EideticNet is easy to implement and train, is efficient, and has time and space complexity linear in the number of parameters. The guarantee of our method holds for normalization layers of modern neural networks during both pre-training and fine-tuning. We show with a variety of network architectures and sets of tasks that EideticNets are immune to forgetting. While the practical benefits of EideticNets are substantial, we believe they can be benefit practitioners and theorists alike. The code for training EideticNets is available at https://github.com/amazon-science/eideticnet-training.

Ahmed Imtiaz Humayun, Randall Balestriero, Richard Baraniuk

Grokking, or delayed generalization, is a phenomenon where generalization in a deep neural network (DNN) occurs long after achieving near zero training error. Previous studies have reported the occurrence of grokking in specific controlled settings, such as DNNs initialized with large-norm parameters or transformers trained on algorithmic datasets. We demonstrate that grokking is actually much more widespread and materializes in a wide range of practical settings, such as training of a convolutional neural network (CNN) on CIFAR10 or a Resnet on Imagenette. We introduce the new concept of delayed robustness, whereby a DNN groks adversarial examples and becomes robust, long after interpolation and/or generalization. We develop an analytical explanation for the emergence of both delayed generalization and delayed robustness based on the local complexity of a DNN's input-output mapping. Our local complexity measures the density of so-called linear regions (aka, spline partition regions) that tile the DNN input space and serves as a utile progress measure for training. We provide the first evidence that, for classification problems, the linear regions undergo a phase transition during training whereafter they migrate away from the training samples (making the DNN mapping smoother there) and towards the decision boundary (making the DNN mapping less smooth there). Grokking occurs post phase transition as a robust partition of the input space thanks to the linearization of the DNN mapping around the training points. Website: https://bit.ly/grok-adversarial

Randall Balestriero, Yann LeCun

Input space reconstruction is an attractive representation learning paradigm. Despite interpretability of the reconstruction and generation, we identify a misalignment between learning by reconstruction, and learning for perception. We show that the former allocates a model's capacity towards a subspace of the data explaining the observed variance--a subspace with uninformative features for the latter. For example, the supervised TinyImagenet task with images projected onto the top subspace explaining 90\% of the pixel variance can be solved with 45\% test accuracy. Using the bottom subspace instead, accounting for only 20\% of the pixel variance, reaches 55\% test accuracy. The features for perception being learned last explains the need for long training time, e.g., with Masked Autoencoders. Learning by denoising is a popular strategy to alleviate that misalignment. We prove that while some noise strategies such as masking are indeed beneficial, others such as additive Gaussian noise are not. Yet, even in the case of masking, we find that the benefits vary as a function of the mask's shape, ratio, and the considered dataset. While tuning the noise strategy without knowledge of the perception task seems challenging, we provide first clues on how to detect if a noise strategy is never beneficial regardless of the perception task.

Patrik Reizinger, Alice Bizeul, Attila Juhos et al.

Supervised learning has become a cornerstone of modern machine learning, yet a comprehensive theory explaining its effectiveness remains elusive. Empirical phenomena, such as neural analogy-making and the linear representation hypothesis, suggest that supervised models can learn interpretable factors of variation in a linear fashion. Recent advances in self-supervised learning, particularly nonlinear Independent Component Analysis, have shown that these methods can recover latent structures by inverting the data generating process. We extend these identifiability results to parametric instance discrimination, then show how insights transfer to the ubiquitous setting of supervised learning with cross-entropy minimization. We prove that even in standard classification tasks, models learn representations of ground-truth factors of variation up to a linear transformation. We corroborate our theoretical contribution with a series of empirical studies. First, using simulated data matching our theoretical assumptions, we demonstrate successful disentanglement of latent factors. Second, we show that on DisLib, a widely-used disentanglement benchmark, simple classification tasks recover latent structures up to linear transformations. Finally, we reveal that models trained on ImageNet encode representations that permit linear decoding of proxy factors of variation. Together, our theoretical findings and experiments offer a compelling explanation for recent observations of linear representations, such as superposition in neural networks. This work takes a significant step toward a cohesive theory that accounts for the unreasonable effectiveness of supervised deep learning.

Polina Kirichenko, Mark Ibrahim, Randall Balestriero et al.

Data augmentation (DA) encodes invariance and provides implicit regularization critical to a model's performance in image classification tasks. However, while DA improves average accuracy, recent studies have shown that its impact can be highly class dependent: achieving optimal average accuracy comes at the cost of significantly hurting individual class accuracy by as much as 20% on ImageNet. There has been little progress in resolving class-level accuracy drops due to a limited understanding of these effects. In this work, we present a framework for understanding how DA interacts with class-level learning dynamics. Using higher-quality multi-label annotations on ImageNet, we systematically categorize the affected classes and find that the majority are inherently ambiguous, co-occur, or involve fine-grained distinctions, while DA controls the model's bias towards one of the closely related classes. While many of the previously reported performance drops are explained by multi-label annotations, our analysis of class confusions reveals other sources of accuracy degradation. We show that simple class-conditional augmentation strategies informed by our framework improve performance on the negatively affected classes.

Vlad Sobal, Wancong Zhang, Kyunghyun Cho et al.

A long-standing goal in AI is to develop agents capable of solving diverse tasks across a range of environments, including those never seen during training. Two dominant paradigms address this challenge: (i) reinforcement learning (RL), which learns policies via trial and error, and (ii) optimal control, which plans actions using a known or learned dynamics model. However, their comparative strengths in the offline setting - where agents must learn from reward-free trajectories - remain underexplored. In this work, we systematically evaluate RL and control-based methods on a suite of navigation tasks, using offline datasets of varying quality. On the RL side, we consider goal-conditioned and zero-shot methods. On the control side, we train a latent dynamics model using the Joint Embedding Predictive Architecture (JEPA) and employ it for planning. We investigate how factors such as data diversity, trajectory quality, and environment variability influence the performance of these approaches. Our results show that model-free RL benefits most from large amounts of high-quality data, whereas model-based planning generalizes better to unseen layouts and is more data-efficient, while achieving trajectory stitching performance comparable to leading model-free methods. Notably, planning with a latent dynamics model proves to be a strong approach for handling suboptimal offline data and adapting to diverse environments.

Hugues Van Assel, Mark Ibrahim, Tommaso Biancalani et al.

Reconstruction and joint embedding have emerged as two leading paradigms in Self Supervised Learning (SSL). Reconstruction methods focus on recovering the original sample from a different view in input space. On the other hand, joint embedding methods align the representations of different views in latent space. Both approaches offer compelling advantages, yet practitioners lack clear guidelines for choosing between them. In this work, we unveil the core mechanisms that distinguish each paradigm. By leveraging closed form solutions for both approaches, we precisely characterize how the view generation process, e.g. data augmentation, impacts the learned representations. We then demonstrate that, unlike supervised learning, both SSL paradigms require a minimal alignment between augmentations and irrelevant features to achieve asymptotic optimality with increasing sample size. Our findings indicate that in scenarios where these irrelevant features have a large magnitude, joint embedding methods are preferable because they impose a strictly weaker alignment condition compared to reconstruction based methods. These results not only clarify the trade offs between the two paradigms but also substantiate the empirical success of joint embedding approaches on real world challenging datasets.

Jakob Ambsdorf, Asbjørn Munk, Sebastian Llambias et al.

With access to large-scale, unlabeled medical datasets, researchers are confronted with two questions: Should they attempt to pretrain a custom foundation model on this medical data, or use transfer-learning from an existing generalist model? And, if a custom model is pretrained, are novel methods required? In this paper we explore these questions by conducting a case-study, in which we train a foundation model on a large regional fetal ultrasound dataset of 2M images. By selecting the well-established DINOv2 method for pretraining, we achieve state-of-the-art results on three fetal ultrasound datasets, covering data from different countries, classification, segmentation, and few-shot tasks. We compare against a series of models pretrained on natural images, ultrasound images, and supervised baselines. Our results demonstrate two key insights: (i) Pretraining on custom data is worth it, even if smaller models are trained on less data, as scaling in natural image pretraining does not translate to ultrasound performance. (ii) Well-tuned methods from computer vision are making it feasible to train custom foundation models for a given medical domain, requiring no hyperparameter tuning and little methodological adaptation. Given these findings, we argue that a bias towards methodological innovation should be avoided when developing domain specific foundation models under common computational resource constraints.

Aarash Feizi, Randall Balestriero, Adriana Romero-Soriano et al.

We propose Guided Positive Sampling Self-Supervised Learning (GPS-SSL), a general method to inject a priori knowledge into Self-Supervised Learning (SSL) positive samples selection. Current SSL methods leverage Data-Augmentations (DA) for generating positive samples and incorporate prior knowledge - an incorrect, or too weak DA will drastically reduce the quality of the learned representation. GPS-SSL proposes instead to design a metric space where Euclidean distances become a meaningful proxy for semantic relationship. In that space, it is now possible to generate positive samples from nearest neighbor sampling. Any prior knowledge can now be embedded into that metric space independently from the employed DA. From its simplicity, GPS-SSL is applicable to any SSL method, e.g. SimCLR or BYOL. A key benefit of GPS-SSL is in reducing the pressure in tailoring strong DAs. For example GPS-SSL reaches 85.58% on Cifar10 with weak DA while the baseline only reaches 37.51%. We therefore move a step forward towards the goal of making SSL less reliant on DA. We also show that even when using strong DAs, GPS-SSL outperforms the baselines on under-studied domains. We evaluate GPS-SSL along with multiple baseline SSL methods on numerous downstream datasets from different domains when the models use strong or minimal data augmentations. We hope that GPS-SSL will open new avenues in studying how to inject a priori knowledge into SSL in a principled manner.

Patrik Reizinger, Randall Balestriero, David Klindt et al.

Self-Supervised Learning (SSL) powers many current AI systems. As research interest and investment grow, the SSL design space continues to expand. The Platonic view of SSL, following the Platonic Representation Hypothesis (PRH), suggests that despite different methods and engineering approaches, all representations converge to the same Platonic ideal. However, this phenomenon lacks precise theoretical explanation. By synthesizing evidence from Identifiability Theory (IT), we show that the PRH can emerge in SSL. However, current IT cannot explain SSL's empirical success. To bridge the gap between theory and practice, we propose expanding IT into what we term Singular Identifiability Theory (SITh), a broader theoretical framework encompassing the entire SSL pipeline. SITh would allow deeper insights into the implicit data assumptions in SSL and advance the field towards learning more interpretable and generalizable representations. We highlight three critical directions for future research: 1) training dynamics and convergence properties of SSL; 2) the impact of finite samples, batch size, and data diversity; and 3) the role of inductive biases in architecture, augmentations, initialization schemes, and optimizers.