Léon Bottou

Alexandre Ramé, Kartik Ahuja, Jianyu Zhang et al.

Foundation models are redefining how AI systems are built. Practitioners now follow a standard procedure to build their machine learning solutions: from a pre-trained foundation model, they fine-tune the weights on the target task of interest. So, the Internet is swarmed by a handful of foundation models fine-tuned on many diverse tasks: these individual fine-tunings exist in isolation without benefiting from each other. In our opinion, this is a missed opportunity, as these specialized models contain rich and diverse features. In this paper, we thus propose model ratatouille, a new strategy to recycle the multiple fine-tunings of the same foundation model on diverse auxiliary tasks. Specifically, we repurpose these auxiliary weights as initializations for multiple parallel fine-tunings on the target task; then, we average all fine-tuned weights to obtain the final model. This recycling strategy aims at maximizing the diversity in weights by leveraging the diversity in auxiliary tasks. Empirically, it improves the state of the art on the reference DomainBed benchmark for out-of-distribution generalization. Looking forward, this work contributes to the emerging paradigm of updatable machine learning where, akin to open-source software development, the community collaborates to reliably update machine learning models. Our code is released: https://github.com/facebookresearch/ModelRatatouille.

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.

Alberto Bietti, Vivien Cabannes, Diane Bouchacourt et al.

Large language models based on transformers have achieved great empirical successes. However, as they are deployed more widely, there is a growing need to better understand their internal mechanisms in order to make them more reliable. These models appear to store vast amounts of knowledge from their training data, and to adapt quickly to new information provided in their context or prompt. We study how transformers balance these two types of knowledge by considering a synthetic setup where tokens are generated from either global or context-specific bigram distributions. By a careful empirical analysis of the training process on a simplified two-layer transformer, we illustrate the fast learning of global bigrams and the slower development of an "induction head" mechanism for the in-context bigrams. We highlight the role of weight matrices as associative memories, provide theoretical insights on how gradients enable their learning during training, and study the role of data-distributional properties.

Jianyu Zhang, David Lopez-Paz, Léon Bottou

There often is a dilemma between ease of optimization and robust out-of-distribution (OoD) generalization. For instance, many OoD methods rely on penalty terms whose optimization is challenging. They are either too strong to optimize reliably or too weak to achieve their goals. We propose to initialize the networks with a rich representation containing a palette of potentially useful features, ready to be used by even simple models. On the one hand, a rich representation provides a good initialization for the optimizer. On the other hand, it also provides an inductive bias that helps OoD generalization. Such a representation is constructed with the Rich Feature Construction (RFC) algorithm, also called the Bonsai algorithm, which consists of a succession of training episodes. During discovery episodes, we craft a multi-objective optimization criterion and its associated datasets in a manner that prevents the network from using the features constructed in the previous iterations. During synthesis episodes, we use knowledge distillation to force the network to simultaneously represent all the previously discovered features. Initializing the networks with Bonsai representations consistently helps six OoD methods achieve top performance on ColoredMNIST benchmark. The same technique substantially outperforms comparable results on the Wilds Camelyon17 task, eliminates the high result variance that plagues other methods, and makes hyperparameter tuning and model selection more reliable.

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.

Jianyu Zhang, Léon Bottou

Does the dominant approach to learn representations (as a side effect of optimizing an expected cost for a single training distribution) remain a good approach when we are dealing with multiple distributions? Our thesis is that such scenarios are better served by representations that are richer than those obtained with a single optimization episode. We support this thesis with simple theoretical arguments and with experiments utilizing an apparently naïve ensembling technique: concatenating the representations obtained from multiple training episodes using the same data, model, algorithm, and hyper-parameters, but different random seeds. These independently trained networks perform similarly. Yet, in a number of scenarios involving new distributions, the concatenated representation performs substantially better than an equivalently sized network trained with a single training run. This proves that the representations constructed by multiple training episodes are in fact different. Although their concatenation carries little additional information about the training task under the training distribution, it becomes substantially more informative when tasks or distributions change. Meanwhile, a single training episode is unlikely to yield such a redundant representation because the optimization process has no reason to accumulate features that do not incrementally improve the training performance.

Zhuoming Chen, Ranajoy Sadhukhan, Zihao Ye et al. · uw

Large language models (LLMs) with long context windows have gained significant attention. However, the KV cache, stored to avoid re-computation, becomes a bottleneck. Various dynamic sparse or TopK-based attention approximation methods have been proposed to leverage the common insight that attention is sparse. In this paper, we first show that TopK attention itself suffers from quality degradation in certain downstream tasks because attention is not always as sparse as expected. Rather than selecting the keys and values with the highest attention scores, sampling with theoretical guarantees can provide a better estimation for attention output. To make the sampling-based approximation practical in LLM generation, we propose MagicPIG, a heterogeneous system based on Locality Sensitive Hashing (LSH). MagicPIG significantly reduces the workload of attention computation while preserving high accuracy for diverse tasks. MagicPIG stores the LSH hash tables and runs the attention computation on the CPU, which allows it to serve longer contexts and larger batch sizes with high approximation accuracy. MagicPIG can improve decoding throughput by up to $5\times$ across various GPU hardware and achieve 54ms decoding latency on a single RTX 4090 for Llama-3.1-8B-Instruct model with a context of 96k tokens. The code is available at https://github.com/Infini-AI-Lab/MagicPIG.

Léon Bottou, Bernhard Schölkopf

Many believe that Large Language Models (LLMs) open the era of Artificial Intelligence (AI). Some see opportunities while others see dangers. Yet both proponents and opponents grasp AI through the imagery popularised by science fiction. Will the machine become sentient and rebel against its creators? Will we experience a paperclip apocalypse? Before answering such questions, we should first ask whether this mental imagery provides a good description of the phenomenon at hand. Understanding weather patterns through the moods of the gods only goes so far. The present paper instead advocates understanding LLMs and their connection to AI through the imagery of Jorge Luis Borges, a master of 20th century literature, forerunner of magical realism, and precursor to postmodern literature. This exercise leads to a new perspective that illuminates the relation between language modelling and artificial intelligence.

Benjamin Aubin, Agnieszka Słowik, Martin Arjovsky et al.

There is an increasing interest in algorithms to learn invariant correlations across training environments. A big share of the current proposals find theoretical support in the causality literature but, how useful are they in practice? The purpose of this note is to propose six linear low-dimensional problems -- unit tests -- to evaluate different types of out-of-distribution generalization in a precise manner. Following initial experiments, none of the three recently proposed alternatives passes all tests. By providing the code to automatically replicate all the results in this manuscript (https://www.github.com/facebookresearch/InvarianceUnitTests), we hope that our unit tests become a standard steppingstone for researchers in out-of-distribution generalization.

Jianyu Zhang, Léon Bottou

It is impossible today to pretend that the practice of machine learning is always compatible with the idea that training and testing data follow the same distribution. Several authors have recently used ensemble techniques to show how scenarios involving multiple data distributions are best served by representations that are both richer than those obtained by regularizing for the best in-distribution performance, and richer than those obtained under the influence of the implicit sparsity bias of common stochastic gradient procedures. This contribution investigates the use of very high dropout rates instead of ensembles to obtain such rich representations. Although training a deep network from scratch using such dropout rates is virtually impossible, fine-tuning a large pre-trained model under such conditions is not only possible but also achieves out-of-distribution performances that exceed those of both ensembles and weight averaging methods such as model soups. This result has practical significance because the importance of the fine-tuning scenario has considerably grown in recent years. This result also provides interesting insights on the nature of rich representations and on the intrinsically linear nature of fine-tuning a large network using a comparatively small dataset.

Jianyu Zhang, Niklas Nolte, Ranajoy Sadhukhan et al.

Memory Mosaics are networks of associative memories working in concert to achieve a prediction task of interest. Like transformers, memory mosaics possess compositional capabilities and in-context learning capabilities. Unlike transformers, memory mosaics achieve these capabilities in comparatively transparent way ("predictive disentanglement"). We illustrate these capabilities on a toy example and also show that memory mosaics perform as well or better than transformers on medium-scale language modeling tasks.

Jianyu Zhang, Léon Bottou

Memory Mosaics [Zhang et al., 2025], networks of associative memories, have demonstrated appealing compositional and in-context learning capabilities on medium-scale networks (GPT-2 scale) and synthetic small datasets. This work shows that these favorable properties remain when we scale memory mosaics to large language model sizes (llama-8B scale) and real-world datasets. To this end, we scale memory mosaics to 10B size, we train them on one trillion tokens, we introduce a couple architectural modifications ("Memory Mosaics v2"), we assess their capabilities across three evaluation dimensions: training-knowledge storage, new-knowledge storage, and in-context learning. Throughout the evaluation, memory mosaics v2 match transformers on the learning of training knowledge (first dimension) and significantly outperforms transformers on carrying out new tasks at inference time (second and third dimensions). These improvements cannot be easily replicated by simply increasing the training data for transformers. A memory mosaics v2 trained on one trillion tokens still perform better on these tasks than a transformer trained on eight trillion tokens.

Jingtong Su, Jianyu Zhang, Karen Ullrich et al.

Common Large Language model (LLM) evaluations rely on demonstration examples to steer models' responses to the desired style. While the number of examples used has been studied and standardized, the choice of how to format examples is less investigated. In evaluation protocols and real world usage, users face the choice how to separate in-context examples: use a comma? new line? semi-colon? hashtag? etc.? Surprisingly, we find this seemingly minor choice can dramatically alter model response quality. Across leading model families (Llama, Qwen, Gemma), performance on MMLU for example can vary by $\pm 23\%$ depending on the choice of delimiter. In fact, one can manipulate model rankings to put any model in the lead by only modifying the single character separating examples. We find LLMs' brittleness pervades topics, model families, and doesn't improve with scale. By probing attention head scores, we find that good-performing delimiters steer attention towards key tokens in the input. Finally, we explore methods to improve LLMs' robustness to the choice of delimiter. We find specifying the selected delimiter in the prompt boosts robustness and offer practical recommendations for the best-performing delimiters to select.

Xingyu Alice Yang, Jianyu Zhang, Léon Bottou

Transfer learning is a cornerstone of modern machine learning, promising a way to adapt models pretrained on a broad mix of data to new tasks with minimal new data. However, a significant challenge remains in ensuring that transferred features are sufficient to handle unseen datasets, amplified by the difficulty of quantifying whether two tasks are "related". To address these challenges, we evaluate model transfer from a pretraining mixture to each of its component tasks, assessing whether pretrained features can match the performance of task-specific direct training. We identify a fundamental limitation in deep learning models -- an "information saturation bottleneck" -- where networks fail to learn new features once they encode similar competing features during training. When restricted to learning only a subset of key features during pretraining, models will permanently lose critical features for transfer and perform inconsistently on data distributions, even components of the training mixture. Empirical evidence from published studies suggests that this phenomenon is pervasive in deep learning architectures -- factors such as data distribution or ordering affect the features that current representation learning methods can learn over time. This study suggests that relying solely on large-scale networks may not be as effective as focusing on task-specific training, when available. We propose richer feature representations as a potential solution to better generalize across new datasets and, specifically, present existing methods alongside a novel approach, the initial steps towards addressing this challenge.

Alexander Peysakhovich, Anna Klimovskaia Susmel, Leon Bottou

Dot product embeddings take a graph and construct vectors for nodes such that dot products between two vectors give the strength of the edge. Dot products make a strong transitivity assumption, however, many important forces generating graphs in the real world lead to non-transitive relationships. We remove the transitivity assumption by embedding nodes into a pseudo-Euclidean space - giving each node an attract and a repel vector. The inner product between two nodes is defined by taking the dot product in attract vectors and subtracting the dot product in repel vectors. Pseudo-Euclidean embeddings can compress networks efficiently, allow for multiple notions of nearest neighbors each with their own interpretation, and can be `slotted' into existing models such as exponential family embeddings or graph neural networks for better link prediction.

Agnieszka Słowik, Léon Bottou

Machine learning systems based on minimizing average error have been shown to perform inconsistently across notable subsets of the data, which is not exposed by a low average error for the entire dataset. In consequential social and economic applications, where data represent people, this can lead to discrimination of underrepresented gender and ethnic groups. Given the importance of bias mitigation in machine learning, the topic leads to contentious debates on how to ensure fairness in practice (data bias versus algorithmic bias). Distributionally Robust Optimization (DRO) seemingly addresses this problem by minimizing the worst expected risk across subpopulations. We establish theoretical results that clarify the relation between DRO and the optimization of the same loss averaged on an adequately weighted training dataset. The results cover finite and infinite number of training distributions, as well as convex and non-convex loss functions. We show that neither DRO nor curating the training set should be construed as a complete solution for bias mitigation: in the same way that there is no universally robust training set, there is no universal way to setup a DRO problem and ensure a socially acceptable set of results. We then leverage these insights to provide a mininal set of practical recommendations for addressing bias with DRO. Finally, we discuss ramifications of our results in other related applications of DRO, using an example of adversarial robustness. Our results show that there is merit to both the algorithm-focused and the data-focused side of the bias debate, as long as arguments in favor of these positions are precisely qualified and backed by relevant mathematics known today.

Alexandre Défossez, Léon Bottou, Francis Bach et al.

We provide a simple proof of convergence covering both the Adam and Adagrad adaptive optimization algorithms when applied to smooth (possibly non-convex) objective functions with bounded gradients. We show that in expectation, the squared norm of the objective gradient averaged over the trajectory has an upper-bound which is explicit in the constants of the problem, parameters of the optimizer, the dimension $d$, and the total number of iterations $N$. This bound can be made arbitrarily small, and with the right hyper-parameters, Adam can be shown to converge with the same rate of convergence $O(d\ln(N)/\sqrt{N})$. When used with the default parameters, Adam doesn't converge, however, and just like constant step-size SGD, it moves away from the initialization point faster than Adagrad, which might explain its practical success. Finally, we obtain the tightest dependency on the heavy ball momentum decay rate $β_1$ among all previous convergence bounds for non-convex Adam and Adagrad, improving from $O((1-β_1)^{-3})$ to $O((1-β_1)^{-1})$.

Alexandre Défossez, Nicolas Usunier, Léon Bottou et al.

Source separation for music is the task of isolating contributions, or stems, from different instruments recorded individually and arranged together to form a song. Such components include voice, bass, drums and any other accompaniments.Contrarily to many audio synthesis tasks where the best performances are achieved by models that directly generate the waveform, the state-of-the-art in source separation for music is to compute masks on the magnitude spectrum. In this paper, we compare two waveform domain architectures. We first adapt Conv-Tasnet, initially developed for speech source separation,to the task of music source separation. While Conv-Tasnet beats many existing spectrogram-domain methods, it suffersfrom significant artifacts, as shown by human evaluations. We propose instead Demucs, a novel waveform-to-waveform model,with a U-Net structure and bidirectional LSTM.Experiments on the MusDB dataset show that, with proper data augmentation, Demucs beats allexisting state-of-the-art architectures, including Conv-Tasnet, with 6.3 SDR on average, (and up to 6.8 with 150 extra training songs, even surpassing the IRM oracle for the bass source).Using recent development in model quantization, Demucs can be compressed down to 120MBwithout any loss of accuracy.We also provide human evaluations, showing that Demucs benefit from a large advantagein terms of the naturalness of the audio. However, it suffers from some bleeding,especially between the vocals and other source.

Zhengdao Chen, Jianyu Zhang, Martin Arjovsky et al.

We propose Symplectic Recurrent Neural Networks (SRNNs) as learning algorithms that capture the dynamics of physical systems from observed trajectories. An SRNN models the Hamiltonian function of the system by a neural network and furthermore leverages symplectic integration, multiple-step training and initial state optimization to address the challenging numerical issues associated with Hamiltonian systems. We show that SRNNs succeed reliably on complex and noisy Hamiltonian systems. We also show how to augment the SRNN integration scheme in order to handle stiff dynamical systems such as bouncing billiards.

Alexandre Défossez, Nicolas Usunier, Léon Bottou et al.

We study the problem of source separation for music using deep learning with four known sources: drums, bass, vocals and other accompaniments. State-of-the-art approaches predict soft masks over mixture spectrograms while methods working on the waveform are lagging behind as measured on the standard MusDB benchmark. Our contribution is two fold. (i) We introduce a simple convolutional and recurrent model that outperforms the state-of-the-art model on waveforms, that is, Wave-U-Net, by 1.6 points of SDR (signal to distortion ratio). (ii) We propose a new scheme to leverage unlabeled music. We train a first model to extract parts with at least one source silent in unlabeled tracks, for instance without bass. We remix this extract with a bass line taken from the supervised dataset to form a new weakly supervised training example. Combining our architecture and scheme, we show that waveform methods can play in the same ballpark as spectrogram ones.

Martin Arjovsky, Léon Bottou, Ishaan Gulrajani et al.

We introduce Invariant Risk Minimization (IRM), a learning paradigm to estimate invariant correlations across multiple training distributions. To achieve this goal, IRM learns a data representation such that the optimal classifier, on top of that data representation, matches for all training distributions. Through theory and experiments, we show how the invariances learned by IRM relate to the causal structures governing the data and enable out-of-distribution generalization.

Aaron Defazio, Léon Bottou

We propose a system for calculating a "scaling constant" for layers and weights of neural networks. We relate this scaling constant to two important quantities that relate to the optimizability of neural networks, and argue that a network that is "preconditioned" via scaling, in the sense that all weights have the same scaling constant, will be easier to train. This scaling calculus results in a number of consequences, among them the fact that the geometric mean of the fan-in and fan-out, rather than the fan-in, fan-out, or arithmetic mean, should be used for the initialization of the variance of weights in a neural network. Our system allows for the off-line design & engineering of ReLU neural networks, potentially replacing blind experimentation.

Chhavi Yadav, Léon Bottou

Although the popular MNIST dataset [LeCun et al., 1994] is derived from the NIST database [Grother and Hanaoka, 1995], the precise processing steps for this derivation have been lost to time. We propose a reconstruction that is accurate enough to serve as a replacement for the MNIST dataset, with insignificant changes in accuracy. We trace each MNIST digit to its NIST source and its rich metadata such as writer identifier, partition identifier, etc. We also reconstruct the complete MNIST test set with 60,000 samples instead of the usual 10,000. Since the balance 50,000 were never distributed, they enable us to investigate the impact of twenty-five years of MNIST experiments on the reported testing performances. Our results unambiguously confirm the trends observed by Recht et al. [2018, 2019]: although the misclassification rates are slightly off, classifier ordering and model selection remain broadly reliable. We attribute this phenomenon to the pairing benefits of comparing classifiers on the same digits.

Aaron Defazio, Léon Bottou

We introduce a new normalization technique that exhibits the fast convergence properties of batch normalization using a transformation of layer weights instead of layer outputs. The proposed technique keeps the contribution of positive and negative weights to the layer output balanced. We validate our method on a set of standard benchmarks including CIFAR-10/100, SVHN and ILSVRC 2012 ImageNet.

Aaron Defazio, Léon Bottou

The application of stochastic variance reduction to optimization has shown remarkable recent theoretical and practical success. The applicability of these techniques to the hard non-convex optimization problems encountered during training of modern deep neural networks is an open problem. We show that naive application of the SVRG technique and related approaches fail, and explore why.

Alexandre Défossez, Neil Zeghidour, Nicolas Usunier et al.

Recent progress in deep learning for audio synthesis opens the way to models that directly produce the waveform, shifting away from the traditional paradigm of relying on vocoders or MIDI synthesizers for speech or music generation. Despite their successes, current state-of-the-art neural audio synthesizers such as WaveNet and SampleRNN suffer from prohibitive training and inference times because they are based on autoregressive models that generate audio samples one at a time at a rate of 16kHz. In this work, we study the more computationally efficient alternative of generating the waveform frame-by-frame with large strides. We present SING, a lightweight neural audio synthesizer for the original task of generating musical notes given desired instrument, pitch and velocity. Our model is trained end-to-end to generate notes from nearly 1000 instruments with a single decoder, thanks to a new loss function that minimizes the distances between the log spectrograms of the generated and target waveforms. On the generalization task of synthesizing notes for pairs of pitch and instrument not seen during training, SING produces audio with significantly improved perceptual quality compared to a state-of-the-art autoencoder based on WaveNet as measured by a Mean Opinion Score (MOS), and is about 32 times faster for training and 2, 500 times faster for inference.

Rachel Ward, Xiaoxia Wu, Leon Bottou

Adaptive gradient methods such as AdaGrad and its variants update the stepsize in stochastic gradient descent on the fly according to the gradients received along the way; such methods have gained widespread use in large-scale optimization for their ability to converge robustly, without the need to fine-tune the stepsize schedule. Yet, the theoretical guarantees to date for AdaGrad are for online and convex optimization. We bridge this gap by providing theoretical guarantees for the convergence of AdaGrad for smooth, nonconvex functions. We show that the norm version of AdaGrad (AdaGrad-Norm) converges to a stationary point at the $\mathcal{O}(\log(N)/\sqrt{N})$ rate in the stochastic setting, and at the optimal $\mathcal{O}(1/N)$ rate in the batch (non-stochastic) setting -- in this sense, our convergence guarantees are 'sharp'. In particular, the convergence of AdaGrad-Norm is robust to the choice of all hyper-parameters of the algorithm, in contrast to stochastic gradient descent whose convergence depends crucially on tuning the step-size to the (generally unknown) Lipschitz smoothness constant and level of stochastic noise on the gradient. Extensive numerical experiments are provided to corroborate our theory; moreover, the experiments suggest that the robustness of AdaGrad-Norm extends to state-of-the-art models in deep learning, without sacrificing generalization.

Xiaoxia Wu, Rachel Ward, Léon Bottou

Adjusting the learning rate schedule in stochastic gradient methods is an important unresolved problem which requires tuning in practice. If certain parameters of the loss function such as smoothness or strong convexity constants are known, theoretical learning rate schedules can be applied. However, in practice, such parameters are not known, and the loss function of interest is not convex in any case. The recently proposed batch normalization reparametrization is widely adopted in most neural network architectures today because, among other advantages, it is robust to the choice of Lipschitz constant of the gradient in loss function, allowing one to set a large learning rate without worry. Inspired by batch normalization, we propose a general nonlinear update rule for the learning rate in batch and stochastic gradient descent so that the learning rate can be initialized at a high value, and is subsequently decreased according to gradient observations along the way. The proposed method is shown to achieve robustness to the relationship between the learning rate and the Lipschitz constant, and near-optimal convergence rates in both the batch and stochastic settings ($O(1/T)$ for smooth loss in the batch setting, and $O(1/\sqrt{T})$ for convex loss in the stochastic setting). We also show through numerical evidence that such robustness of the proposed method extends to highly nonconvex and possibly non-smooth loss function in deep learning problems.Our analysis establishes some first theoretical understanding into the observed robustness for batch normalization and weight normalization.

Carl-Johann Simon-Gabriel, Yann Ollivier, Léon Bottou et al.

Over the past few years, neural networks were proven vulnerable to adversarial images: targeted but imperceptible image perturbations lead to drastically different predictions. We show that adversarial vulnerability increases with the gradients of the training objective when viewed as a function of the inputs. Surprisingly, vulnerability does not depend on network topology: for many standard network architectures, we prove that at initialization, the $\ell_1$-norm of these gradients grows as the square root of the input dimension, leaving the networks increasingly vulnerable with growing image size. We empirically show that this dimension dependence persists after either usual or robust training, but gets attenuated with higher regularization.

Leon Bottou, Martin Arjovsky, David Lopez-Paz et al.

Learning algorithms for implicit generative models can optimize a variety of criteria that measure how the data distribution differs from the implicit model distribution, including the Wasserstein distance, the Energy distance, and the Maximum Mean Discrepancy criterion. A careful look at the geometries induced by these distances on the space of probability measures reveals interesting differences. In particular, we can establish surprising approximate global convergence guarantees for the $1$-Wasserstein distance,even when the parametric generator has a nonconvex parametrization.

Levent Sagun, Utku Evci, V. Ugur Guney et al.

We study the properties of common loss surfaces through their Hessian matrix. In particular, in the context of deep learning, we empirically show that the spectrum of the Hessian is composed of two parts: (1) the bulk centered near zero, (2) and outliers away from the bulk. We present numerical evidence and mathematical justifications to the following conjectures laid out by Sagun et al. (2016): Fixing data, increasing the number of parameters merely scales the bulk of the spectrum; fixing the dimension and changing the data (for instance adding more clusters or making the data less separable) only affects the outliers. We believe that our observations have striking implications for non-convex optimization in high dimensions. First, the flatness of such landscapes (which can be measured by the singularity of the Hessian) implies that classical notions of basins of attraction may be quite misleading. And that the discussion of wide/narrow basins may be in need of a new perspective around over-parametrization and redundancy that are able to create large connected components at the bottom of the landscape. Second, the dependence of small number of large eigenvalues to the data distribution can be linked to the spectrum of the covariance matrix of gradients of model outputs. With this in mind, we may reevaluate the connections within the data-architecture-algorithm framework of a model, hoping that it would shed light into the geometry of high-dimensional and non-convex spaces in modern applications. In particular, we present a case that links the two observations: small and large batch gradient descent appear to converge to different basins of attraction but we show that they are in fact connected through their flat region and so belong to the same basin.

Jean Lafond, Nicolas Vasilache, Léon Bottou

We define a second-order neural network stochastic gradient training algorithm whose block-diagonal structure effectively amounts to normalizing the unit activations. Investigating why this algorithm lacks in robustness then reveals two interesting insights. The first insight suggests a new way to scale the stepsizes, clarifying popular algorithms such as RMSProp as well as old neural network tricks such as fanin stepsize scaling. The second insight stresses the practical importance of dealing with fast changes of the curvature of the cost.

Martin Arjovsky, Soumith Chintala, Léon Bottou

We introduce a new algorithm named WGAN, an alternative to traditional GAN training. In this new model, we show that we can improve the stability of learning, get rid of problems like mode collapse, and provide meaningful learning curves useful for debugging and hyperparameter searches. Furthermore, we show that the corresponding optimization problem is sound, and provide extensive theoretical work highlighting the deep connections to other distances between distributions.

Martin Arjovsky, Léon Bottou

The goal of this paper is not to introduce a single algorithm or method, but to make theoretical steps towards fully understanding the training dynamics of generative adversarial networks. In order to substantiate our theoretical analysis, we perform targeted experiments to verify our assumptions, illustrate our claims, and quantify the phenomena. This paper is divided into three sections. The first section introduces the problem at hand. The second section is dedicated to studying and proving rigorously the problems including instability and saturation that arize when training generative adversarial networks. The third section examines a practical and theoretically grounded direction towards solving these problems, while introducing new tools to study them.

Levent Sagun, Leon Bottou, Yann LeCun

We look at the eigenvalues of the Hessian of a loss function before and after training. The eigenvalue distribution is seen to be composed of two parts, the bulk which is concentrated around zero, and the edges which are scattered away from zero. We present empirical evidence for the bulk indicating how over-parametrized the system is, and for the edges that depend on the input data.

Léon Bottou, Frank E. Curtis, Jorge Nocedal

This paper provides a review and commentary on the past, present, and future of numerical optimization algorithms in the context of machine learning applications. Through case studies on text classification and the training of deep neural networks, we discuss how optimization problems arise in machine learning and what makes them challenging. A major theme of our study is that large-scale machine learning represents a distinctive setting in which the stochastic gradient (SG) method has traditionally played a central role while conventional gradient-based nonlinear optimization techniques typically falter. Based on this viewpoint, we present a comprehensive theory of a straightforward, yet versatile SG algorithm, discuss its practical behavior, and highlight opportunities for designing algorithms with improved performance. This leads to a discussion about the next generation of optimization methods for large-scale machine learning, including an investigation of two main streams of research on techniques that diminish noise in the stochastic directions and methods that make use of second-order derivative approximations.

David Lopez-Paz, Robert Nishihara, Soumith Chintala et al.

This paper establishes the existence of observable footprints that reveal the "causal dispositions" of the object categories appearing in collections of images. We achieve this goal in two steps. First, we take a learning approach to observational causal discovery, and build a classifier that achieves state-of-the-art performance on finding the causal direction between pairs of random variables, given samples from their joint distribution. Second, we use our causal direction classifier to effectively distinguish between features of objects and features of their contexts in collections of static images. Our experiments demonstrate the existence of a relation between the direction of causality and the difference between objects and their contexts, and by the same token, the existence of observable signals that reveal the causal dispositions of objects.

David Lopez-Paz, Léon Bottou, Bernhard Schölkopf et al.

Distillation (Hinton et al., 2015) and privileged information (Vapnik & Izmailov, 2015) are two techniques that enable machines to learn from other machines. This paper unifies these two techniques into generalized distillation, a framework to learn from multiple machines and data representations. We provide theoretical and causal insight about the inner workings of generalized distillation, extend it to unsupervised, semisupervised and multitask learning scenarios, and illustrate its efficacy on a variety of numerical simulations on both synthetic and real-world data.

Robert Nishihara, David Lopez-Paz, Léon Bottou

Algorithms for hyperparameter optimization abound, all of which work well under different and often unverifiable assumptions. Motivated by the general challenge of sequentially choosing which algorithm to use, we study the more specific task of choosing among distributions to use for random hyperparameter optimization. This work is naturally framed in the extreme bandit setting, which deals with sequentially choosing which distribution from a collection to sample in order to minimize (maximize) the single best cost (reward). Whereas the distributions in the standard bandit setting are primarily characterized by their means, a number of subtleties arise when we care about the minimal cost as opposed to the average cost. For example, there may not be a well-defined "best" distribution as there is in the standard bandit setting. The best distribution depends on the rewards that have been obtained and on the remaining time horizon. Whereas in the standard bandit setting, it is sensible to compare policies with an oracle which plays the single best arm, in the extreme bandit setting, there are multiple sensible oracle models. We define a sensible notion of "extreme regret" in the extreme bandit setting, which parallels the concept of regret in the standard bandit setting. We then prove that no policy can asymptotically achieve no extreme regret.

Alekh Agarwal, Leon Bottou

This paper presents a lower bound for optimizing a finite sum of $n$ functions, where each function is $L$-smooth and the sum is $μ$-strongly convex. We show that no algorithm can reach an error $ε$ in minimizing all functions from this class in fewer than $Ω(n + \sqrt{n(κ-1)}\log(1/ε))$ iterations, where $κ=L/μ$ is a surrogate condition number. We then compare this lower bound to upper bounds for recently developed methods specializing to this setting. When the functions involved in this sum are not arbitrary, but based on i.i.d. random data, then we further contrast these complexity results with those for optimal first-order methods to directly optimize the sum. The conclusion we draw is that a lot of caution is necessary for an accurate comparison, and identify machine learning scenarios where the new methods help computationally.

Patrice Simard, David Chickering, Aparna Lakshmiratan et al.

Quick interaction between a human teacher and a learning machine presents numerous benefits and challenges when working with web-scale data. The human teacher guides the machine towards accomplishing the task of interest. The learning machine leverages big data to find examples that maximize the training value of its interaction with the teacher. When the teacher is restricted to labeling examples selected by the machine, this problem is an instance of active learning. When the teacher can provide additional information to the machine (e.g., suggestions on what examples or predictive features should be used) as the learning task progresses, then the problem becomes one of interactive learning. To accommodate the two-way communication channel needed for efficient interactive learning, the teacher and the machine need an environment that supports an interaction language. The machine can access, process, and summarize more examples than the teacher can see in a lifetime. Based on the machine's output, the teacher can revise the definition of the task or make it more precise. Both the teacher and the machine continuously learn and benefit from the interaction. We have built a platform to (1) produce valuable and deployable models and (2) support research on both the machine learning and user interface challenges of the interactive learning problem. The platform relies on a dedicated, low-latency, distributed, in-memory architecture that allows us to construct web-scale learning machines with quick interaction speed. The purpose of this paper is to describe this architecture and demonstrate how it supports our research efforts. Preliminary results are presented as illustrations of the architecture but are not the primary focus of the paper.

Dhruv Mahajan, S. Sathiya Keerthi, S. Sundararajan et al.

This paper proposes a novel parallel stochastic gradient descent (SGD) method that is obtained by applying parallel sets of SGD iterations (each set operating on one node using the data residing in it) for finding the direction in each iteration of a batch descent method. The method has strong convergence properties. Experiments on datasets with high dimensional feature spaces show the value of this method.

Dhruv Mahajan, Nikunj Agrawal, S. Sathiya Keerthi et al.

Scalable machine learning over big data is an important problem that is receiving a lot of attention in recent years. On popular distributed environments such as Hadoop running on a cluster of commodity machines, communication costs are substantial and algorithms need to be designed suitably considering those costs. In this paper we give a novel approach to the distributed training of linear classifiers (involving smooth losses and L2 regularization) that is designed to reduce the total communication costs. At each iteration, the nodes minimize locally formed approximate objective functions; then the resulting minimizers are combined to form a descent direction to move. Our approach gives a lot of freedom in the formation of the approximate objective function as well as in the choice of methods to solve them. The method is shown to have $O(log(1/ε))$ time convergence. The method can be viewed as an iterative parameter mixing method. A special instantiation yields a parallel stochastic gradient descent method with strong convergence. When communication times between nodes are large, our method is much faster than the Terascale method (Agarwal et al., 2011), which is a state of the art distributed solver based on the statistical query model (Chuet al., 2006) that computes function and gradient values in a distributed fashion. We also evaluate against other recent distributed methods and demonstrate superior performance of our method.

Alekh Agarwal, Leon Bottou, Miroslav Dudik et al.

Training examples are not all equally informative. Active learning strategies leverage this observation in order to massively reduce the number of examples that need to be labeled. We leverage the same observation to build a generic strategy for parallelizing learning algorithms. This strategy is effective because the search for informative examples is highly parallelizable and because we show that its performance does not deteriorate when the sifting process relies on a slightly outdated model. Parallel active learning is particularly attractive to train nonlinear models with non-linear representations because there are few practical parallel learning algorithms for such models. We report preliminary experiments using both kernel SVMs and SGD-trained neural networks.

Léon Bottou, Jonas Peters, Joaquin Quiñonero-Candela et al.

This work shows how to leverage causal inference to understand the behavior of complex learning systems interacting with their environment and predict the consequences of changes to the system. Such predictions allow both humans and algorithms to select changes that improve both the short-term and long-term performance of such systems. This work is illustrated by experiments carried out on the ad placement system associated with the Bing search engine.