arXiv subjects

Andrea Montanari

Publications and source records attributed to Andrea Montanari.

At least 55 records · Page 3Linked to original sources

arXiv2023

Six Lectures on Linearized Neural Networks

In these six lectures, we examine what can be learnt about the behavior of multi-layer neural networks from the analysis of linear models. We first recall the correspondence between neural networks and linear models via the so-called lazy regime. We then review four models for linearized neural networks: linear regression with concentrated features, kernel ridge regression, random feature model and neural tangent model. Finally, we highlight the limitations of the linear theory and discuss how other approaches can overcome them.

stat.ML↗

arXiv2023

Tractability from overparametrization: The example of the negative perceptron

In the negative perceptron problem we are given $n$ data points $({\boldsymbol x}_i,y_i)$, where ${\boldsymbol x}_i$ is a $d$-dimensional vector and $y_i\in\{+1,-1\}$ is a binary label. The data are not linearly separable and hence we content ourselves to find a linear classifier with the largest possible \emph{negative} margin. In other words, we want to find a unit norm vector ${\boldsymbol θ}$ that maximizes $\min_{i\le n}y_i\langle {\boldsymbol θ},{\boldsymbol x}_i\rangle$. This is a non-convex optimization problem (it is equivalent to finding a maximum norm vector in a polytope), and we study its typical properties under two random models for the data. We consider the proportional asymptotics in which $n,d\to \infty$ with $n/d\toδ$, and prove upper and lower bounds on the maximum margin $κ_{\text{s}}(δ)$ or -- equivalently -- on its inverse function $δ_{\text{s}}(κ)$. In other words, $δ_{\text{s}}(κ)$ is the overparametrization threshold: for $n/d\le δ_{\text{s}}(κ)-\varepsilon$ a classifier achieving vanishing training error exists with high probability, while for $n/d\ge δ_{\text{s}}(κ)+\varepsilon$ it does not. Our bounds on $δ_{\text{s}}(κ)$ match to the leading order as $κ\to -\infty$. We then analyze a linear programming algorithm to find a solution, and characterize the corresponding threshold $δ_{\text{lin}}(κ)$. We observe a gap between the interpolation threshold $δ_{\text{s}}(κ)$ and the linear programming threshold $δ_{\text{lin}}(κ)$, raising the question of the behavior of other algorithms.

cs.LG↗

arXiv2023

The generalization error of max-margin linear classifiers: Benign overfitting and high dimensional asymptotics in the overparametrized regime

Modern machine learning classifiers often exhibit vanishing classification error on the training set. They achieve this by learning nonlinear representations of the inputs that maps the data into linearly separable classes. Motivated by these phenomena, we revisit high-dimensional maximum margin classification for linearly separable data. We consider a stylized setting in which data $(y_i,{\boldsymbol x}_i)$, $i\le n$ are i.i.d. with ${\boldsymbol x}_i\sim\mathsf{N}({\boldsymbol 0},{\boldsymbol Σ})$ a $p$-dimensional Gaussian feature vector, and $y_i \in\{+1,-1\}$ a label whose distribution depends on a linear combination of the covariates $\langle {\boldsymbol θ}_*,{\boldsymbol x}_i \rangle$. While the Gaussian model might appear extremely simplistic, universality arguments can be used to show that the results derived in this setting also apply to the output of certain nonlinear featurization maps. We consider the proportional asymptotics $n,p\to\infty$ with $p/n\to ψ$, and derive exact expressions for the limiting generalization error. We use this theory to derive two results of independent interest: $(i)$ Sufficient conditions on $({\boldsymbol Σ},{\boldsymbol θ}_*)$ for `benign overfitting' that parallel previously derived conditions in the case of linear regression; $(ii)$ An asymptotically exact expression for the generalization error when max-margin classification is used in conjunction with feature vectors produced by random one-layer neural networks.

math.ST↗

arXiv2023

Compressing Tabular Data via Latent Variable Estimation

Data used for analytics and machine learning often take the form of tables with categorical entries. We introduce a family of lossless compression algorithms for such data that proceed in four steps: $(i)$ Estimate latent variables associated to rows and columns; $(ii)$ Partition the table in blocks according to the row/column latents; $(iii)$ Apply a sequential (e.g. Lempel-Ziv) coder to each of the blocks; $(iv)$ Append a compressed encoding of the latents. We evaluate it on several benchmark datasets, and study optimal compression in a probabilistic model for that tabular data, whereby latent values are independent and table entries are conditionally independent given the latent values. We prove that the model has a well defined entropy rate and satisfies an asymptotic equipartition property. We also prove that classical compression schemes such as Lempel-Ziv and finite-state encoders do not achieve this rate. On the other hand, the latent estimation strategy outlined above achieves the optimal rate.

cs.IT↗

arXiv2023

Local algorithms for Maximum Cut and Minimum Bisection on locally treelike regular graphs of large degree

Given a graph $G$ of degree $k$ over $n$ vertices, we consider the problem of computing a near maximum cut or a near minimum bisection in polynomial time. For graphs of girth $2L$, we develop a local message passing algorithm whose complexity is $O(nkL)$, and that achieves near optimal cut values among all $L$-local algorithms. Focusing on max-cut, the algorithm constructs a cut of value $nk/4+ n\mathsf{P}_\star\sqrt{k/4}+\mathsf{err}(n,k,L)$, where $\mathsf{P}_\star\approx 0.763166$ is the value of the Parisi formula from spin glass theory, and $\mathsf{err}(n,k,L)=o_n(n)+no_k(\sqrt{k})+n \sqrt{k} o_L(1)$ (subscripts indicate the asymptotic variables). Our result generalizes to locally treelike graphs, i.e., graphs whose girth becomes $2L$ after removing a small fraction of vertices. Earlier work established that, for random $k$-regular graphs, the typical max-cut value is $nk/4+ n\mathsf{P}_\star\sqrt{k/4}+o_n(n)+no_k(\sqrt{k})$. Therefore our algorithm is nearly optimal on such graphs. An immediate corollary of this result is that random regular graphs have nearly minimum max-cut, and nearly maximum min-bisection among all regular locally treelike graphs. This can be viewed as a combinatorial version of the near-Ramanujan property of random regular graphs.

math.PR↗

arXiv2023

Adversarial Examples in Random Neural Networks with General Activations

A substantial body of empirical work documents the lack of robustness in deep learning models to adversarial examples. Recent theoretical work proved that adversarial examples are ubiquitous in two-layers networks with sub-exponential width and ReLU or smooth activations, and multi-layer ReLU networks with sub-exponential width. We present a result of the same type, with no restriction on width and for general locally Lipschitz continuous activations. More precisely, given a neural network $f(\,\cdot\,;{\boldsymbol θ})$ with random weights ${\boldsymbol θ}$, and feature vector ${\boldsymbol x}$, we show that an adversarial example ${\boldsymbol x}'$ can be found with high probability along the direction of the gradient $\nabla_{\boldsymbol x}f({\boldsymbol x};{\boldsymbol θ})$. Our proof is based on a Gaussian conditioning technique. Instead of proving that $f$ is approximately linear in a neighborhood of ${\boldsymbol x}$, we characterize the joint distribution of $f({\boldsymbol x};{\boldsymbol θ})$ and $f({\boldsymbol x}';{\boldsymbol θ})$ for ${\boldsymbol x}' = {\boldsymbol x}-s({\boldsymbol x})\nabla_{\boldsymbol x}f({\boldsymbol x};{\boldsymbol θ})$.

cs.LG↗

arXiv2022

Fundamental Limits of Low-Rank Matrix Estimation with Diverging Aspect Ratios

We consider the problem of estimating the factors of a low-rank $n \times d$ matrix, when this is corrupted by additive Gaussian noise. A special example of our setting corresponds to clustering mixtures of Gaussians with equal (known) covariances. Simple spectral methods do not take into account the distribution of the entries of these factors and are therefore often suboptimal. Here, we characterize the asymptotics of the minimum estimation error under the assumption that the distribution of the entries is known to the statistician. Our results apply to the high-dimensional regime $n, d \to \infty$ and $d / n \to \infty$ (or $d / n \to 0$) and generalize earlier work that focused on the proportional asymptotics $n, d \to \infty$, $d / n \to δ\in (0, \infty)$. We outline an interesting signal strength regime in which $d / n \to \infty$ and partial recovery is possible for the left singular vectors while impossible for the right singular vectors. We illustrate the general theory by deriving consequences for Gaussian mixture clustering and carrying out a numerical study on genomics data.

math.ST↗

arXiv2022

Optimization of random high-dimensional functions: Structure and algorithms

Replica symmetry breaking postulates that near optima of spin glass Hamiltonians have an ultrametric structure. Namely, near optima can be associated to leaves of a tree, and the Euclidean distance between them corresponds to the distance along this tree. We survey recent progress towards a rigorous proof of this picture in the context of mixed $p$-spin spin glass models. We focus in particular on the following topics: $(i)$~The structure of critical points of the Hamiltonian; $(ii)$~The realization of the ultrametric tree as near optima of a suitable TAP free energy; $(iii)$~The construction of efficient optimization algorithm that exploits this picture.

math.PR↗

arXiv2022

The Interpolation Phase Transition in Neural Networks: Memorization and Generalization under Lazy Training

Modern neural networks are often operated in a strongly overparametrized regime: they comprise so many parameters that they can interpolate the training set, even if actual labels are replaced by purely random ones. Despite this, they achieve good prediction error on unseen data: interpolating the training set does not lead to a large generalization error. Further, overparametrization appears to be beneficial in that it simplifies the optimization landscape. Here we study these phenomena in the context of two-layers neural networks in the neural tangent (NT) regime. We consider a simple data model, with isotropic covariates vectors in $d$ dimensions, and $N$ hidden neurons. We assume that both the sample size $n$ and the dimension $d$ are large, and they are polynomially related. Our first main result is a characterization of the eigenstructure of the empirical NT kernel in the overparametrized regime $Nd\gg n$. This characterization implies as a corollary that the minimum eigenvalue of the empirical NT kernel is bounded away from zero as soon as $Nd\gg n$, and therefore the network can exactly interpolate arbitrary labels in the same regime. Our second main result is a characterization of the generalization error of NT ridge regression including, as a special case, min-$\ell_2$ norm interpolation. We prove that, as soon as $Nd\gg n$, the test error is well approximated by the one of kernel ridge regression with respect to the infinite-width kernel. The latter is in turn well approximated by the error of polynomial ridge regression, whereby the regularization parameter is increased by a `self-induced' term related to the high-degree components of the activation function. The polynomial degree depends on the sample size and the dimension (in particular on $\log n/\log d$).

stat.ML↗

arXiv2022

Fundamental Barriers to High-Dimensional Regression with Convex Penalties

In high-dimensional regression, we attempt to estimate a parameter vector $β_0\in\mathbb{R}^p$ from $n\lesssim p$ observations $\{(y_i,x_i)\}_{i\leq n}$ where $x_i\in\mathbb{R}^p$ is a vector of predictors and $y_i$ is a response variable. A well-established approach uses convex regularizers to promote specific structures (e.g. sparsity) of the estimate $\widehatβ$, while allowing for practical algorithms. Theoretical analysis implies that convex penalization schemes have nearly optimal estimation properties in certain settings. However, in general the gaps between statistically optimal estimation (with unbounded computational resources) and convex methods are poorly understood. We show that when the statistican has very simple structural information about the distribution of the entries of $β_0$, a large gap frequently exists between the best performance achieved by any convex regularizer satisfying a mild technical condition and either (i) the optimal statistical error or (ii) the statistical error achieved by optimal approximate message passing algorithms. Remarkably, a gap occurs at high enough signal-to-noise ratio if and only if the distribution of the coordinates of $β_0$ is not log-concave. These conclusions follow from an analysis of standard Gaussian designs. Our lower bounds are expected to be generally tight, and we prove tightness under certain conditions.

math.ST↗

arXiv2022

Statistically Optimal First Order Algorithms: A Proof via Orthogonalization

We consider a class of statistical estimation problems in which we are given a random data matrix ${\boldsymbol X}\in {\mathbb R}^{n\times d}$ (and possibly some labels ${\boldsymbol y}\in{\mathbb R}^n$) and would like to estimate a coefficient vector ${\boldsymbol θ}\in{\mathbb R}^d$ (or possibly a constant number of such vectors). Special cases include low-rank matrix estimation and regularized estimation in generalized linear models (e.g., sparse regression). First order methods proceed by iteratively multiplying current estimates by ${\boldsymbol X}$ or its transpose. Examples include gradient descent or its accelerated variants. Celentano, Montanari, Wu proved that for any constant number of iterations (matrix vector multiplications), the optimal first order algorithm is a specific approximate message passing algorithm (known as `Bayes AMP'). The error of this estimator can be characterized in the high-dimensional asymptotics $n,d\to\infty$, $n/d\toδ$, and provides a lower bound to the estimation error of any first order algorithm. Here we present a simpler proof of the same result, and generalize it to broader classes of data distributions and of first order algorithms, including algorithms with non-separable nonlinearities. Most importantly, the new proof technique does not require to construct an equivalent tree-structured estimation problem, and is therefore susceptible of a broader range of applications.

stat.CO↗

arXiv2021

When Do Neural Networks Outperform Kernel Methods?

For a certain scaling of the initialization of stochastic gradient descent (SGD), wide neural networks (NN) have been shown to be well approximated by reproducing kernel Hilbert space (RKHS) methods. Recent empirical work showed that, for some classification tasks, RKHS methods can replace NNs without a large loss in performance. On the other hand, two-layers NNs are known to encode richer smoothness classes than RKHS and we know of special examples for which SGD-trained NN provably outperform RKHS. This is true even in the wide network limit, for a different scaling of the initialization. How can we reconcile the above claims? For which tasks do NNs outperform RKHS? If covariates are nearly isotropic, RKHS methods suffer from the curse of dimensionality, while NNs can overcome it by learning the best low-dimensional representation. Here we show that this curse of dimensionality becomes milder if the covariates display the same low-dimensional structure as the target function, and we precisely characterize this tradeoff. Building on these results, we present the spiked covariates model that can capture in a unified framework both behaviors observed in earlier work. We hypothesize that such a latent low-dimensional structure is present in image classification. We test numerically this hypothesis by showing that specific perturbations of the training distribution degrade the performances of RKHS methods much more significantly than NNs.

stat.ML↗

arXiv2021

Minimum complexity interpolation in random features models

Despite their many appealing properties, kernel methods are heavily affected by the curse of dimensionality. For instance, in the case of inner product kernels in $\mathbb{R}^d$, the Reproducing Kernel Hilbert Space (RKHS) norm is often very large for functions that depend strongly on a small subset of directions (ridge functions). Correspondingly, such functions are difficult to learn using kernel methods. This observation has motivated the study of generalizations of kernel methods, whereby the RKHS norm -- which is equivalent to a weighted $\ell_2$ norm -- is replaced by a weighted functional $\ell_p$ norm, which we refer to as $\mathcal{F}_p$ norm. Unfortunately, tractability of these approaches is unclear. The kernel trick is not available and minimizing these norms requires to solve an infinite-dimensional convex problem. We study random features approximations to these norms and show that, for $p>1$, the number of random features required to approximate the original learning problem is upper bounded by a polynomial in the sample size. Hence, learning with $\mathcal{F}_p$ norms is tractable in these cases. We introduce a proof technique based on uniform concentration in the dual, which can be of broader interest in the study of overparametrized models. For $p= 1$, our guarantees for the random features approximation break down. We prove instead that learning with the $\mathcal{F}_1$ norm is $\mathsf{NP}$-hard under a randomized reduction based on the problem of learning halfspaces with noise.

cs.LG↗

arXiv2021

An Information-Theoretic View of Stochastic Localization

Given a probability measure $μ$ over ${\mathbb R}^n$, it is often useful to approximate it by the convex combination of a small number of probability measures, such that each component is close to a product measure. Recently, Ronen Eldan used a stochastic localization argument to prove a general decomposition result of this type. In Eldan's theorem, the `number of components' is characterized by the entropy of the mixture, and `closeness to product' is characterized by the covariance matrix of each component. We present an elementary proof of Eldan's theorem which makes use of an information theory (or estimation theory) interpretation. The proof is analogous to the one of an earlier decomposition result known as the `pinning lemma.'

cs.IT↗

arXiv2021

CAD: Debiasing the Lasso with inaccurate covariate model

We consider the problem of estimating a low-dimensional parameter in high-dimensional linear regression. Constructing an approximately unbiased estimate of the parameter of interest is a crucial step towards performing statistical inference. Several authors suggest to orthogonalize both the variable of interest and the outcome with respect to the nuisance variables, and then regress the residual outcome with respect to the residual variable. This is possible if the covariance structure of the regressors is perfectly known, or is sufficiently structured that it can be estimated accurately from data (e.g., the precision matrix is sufficiently sparse). Here we consider a regime in which the covariate model can only be estimated inaccurately, and hence existing debiasing approaches are not guaranteed to work. When errors in estimating the covariate model are correlated with errors in estimating the linear model parameter, an incomplete elimination of the bias occurs. We propose the Correlation Adjusted Debiased Lasso (CAD), which nearly eliminates this bias in some cases, including cases in which the estimation errors are neither negligible nor orthogonal. We consider a setting in which some unlabeled samples might be available to the statistician alongside labeled ones (semi-supervised learning), and our guarantees hold under the assumption of jointly Gaussian covariates. The new debiased estimator is guaranteed to cancel the bias in two cases: (1) when the total number of samples (labeled and unlabeled) is larger than the number of parameters, or (2) when the covariance of the nuisance (but not the effect of the nuisance on the variable of interest) is known. Neither of these cases is treated by state-of-the-art methods.

math.ST↗

arXiv2021

Streaming Belief Propagation for Community Detection

The community detection problem requires to cluster the nodes of a network into a small number of well-connected "communities". There has been substantial recent progress in characterizing the fundamental statistical limits of community detection under simple stochastic block models. However, in real-world applications, the network structure is typically dynamic, with nodes that join over time. In this setting, we would like a detection algorithm to perform only a limited number of updates at each node arrival. While standard voting approaches satisfy this constraint, it is unclear whether they exploit the network information optimally. We introduce a simple model for networks growing over time which we refer to as streaming stochastic block model (StSBM). Within this model, we prove that voting algorithms have fundamental limitations. We also develop a streaming belief-propagation (StreamBP) approach, for which we prove optimality in certain regimes. We validate our theoretical findings on synthetic and real data.

stat.ML↗

arXiv2021

Deep learning: a statistical viewpoint

The remarkable practical success of deep learning has revealed some major surprises from a theoretical perspective. In particular, simple gradient methods easily find near-optimal solutions to non-convex optimization problems, and despite giving a near-perfect fit to training data without any explicit effort to control model complexity, these methods exhibit excellent predictive accuracy. We conjecture that specific principles underlie these phenomena: that overparametrization allows gradient methods to find interpolating solutions, that these methods implicitly impose regularization, and that overparametrization leads to benign overfitting. We survey recent theoretical progress that provides examples illustrating these principles in simpler settings. We first review classical uniform convergence results and why they fall short of explaining aspects of the behavior of deep learning methods. We give examples of implicit regularization in simple settings, where gradient methods lead to minimal norm functions that perfectly fit the training data. Then we review prediction methods that exhibit benign overfitting, focusing on regression problems with quadratic loss. For these methods, we can decompose the prediction rule into a simple component that is useful for prediction and a spiky component that is useful for overfitting but, in a favorable setting, does not harm prediction accuracy. We focus specifically on the linear regime for neural networks, where the network can be approximated by a linear model. In this regime, we demonstrate the success of gradient flow, and we consider benign overfitting with two-layer networks, giving an exact asymptotic analysis that precisely demonstrates the impact of overparametrization. We conclude by highlighting the key challenges that arise in extending these insights to realistic deep learning settings.

math.ST↗

arXiv2021

Learning with invariances in random features and kernel models

A number of machine learning tasks entail a high degree of invariance: the data distribution does not change if we act on the data with a certain group of transformations. For instance, labels of images are invariant under translations of the images. Certain neural network architectures -- for instance, convolutional networks -- are believed to owe their success to the fact that they exploit such invariance properties. With the objective of quantifying the gain achieved by invariant architectures, we introduce two classes of models: invariant random features and invariant kernel methods. The latter includes, as a special case, the neural tangent kernel for convolutional networks with global average pooling. We consider uniform covariates distributions on the sphere and hypercube and a general invariant target function. We characterize the test error of invariant methods in a high-dimensional regime in which the sample size and number of hidden units scale as polynomials in the dimension, for a class of groups that we call `degeneracy $α$', with $α\leq 1$. We show that exploiting invariance in the architecture saves a $d^α$ factor ($d$ stands for the dimension) in sample size and number of hidden units to achieve the same test error as for unstructured architectures. Finally, we show that output symmetrization of an unstructured kernel estimator does not give a significant statistical improvement; on the other hand, data augmentation with an unstructured kernel estimator is equivalent to an invariant kernel estimator and enjoys the same improvement in statistical efficiency.

stat.ML↗