论文检索

Deep Learning/Theory

Sébastien Lachapelle, Divyat Mahajan, Ioannis Mitliagkas, Simon Lacoste-Julien

We tackle the problems of latent variables identification and "out-of-support'' image generation in representation learning. We show that both are possible for a class of decoders that we call additive, which are reminiscent of decoders used for object-centric representation learning (OCRL) and well suited for images that can be decomposed as a sum of object-specific images. We provide conditions under which exactly solving the reconstruction problem using an additive decoder is guaranteed to identify the blocks of latent variables up to permutation and block-wise invertible transformations. This guarantee relies only on very weak assumptions about the distribution of the latent factors, which might present statistical dependencies and have an almost arbitrarily shaped support. Our result provides a new setting where nonlinear independent component analysis (ICA) is possible and adds to our theoretical understanding of OCRL methods. We also show theoretically that additive decoders can generate novel images by recombining observed factors of variations in novel ways, an ability we refer to as Cartesian-product extrapolation. We show empirically that additivity is crucial for both identifiability and extrapolation on simulated data.

Deep Learning/Theory

Lachlan MacDonald, Jack Valmadre, Hemanth Saratchandran, Simon Lucey

We introduce a general theoretical framework, designed for the study of gradient optimisation of deep neural networks, that encompasses ubiquitous architecture choices including batch normalisation, weight normalisation and skip connections. Our framework determines the curvature and regularity properties of multilayer loss landscapes in terms of their constituent layers, thereby elucidating the roles played by normalisation layers and skip connections in globalising these properties. We then demonstrate the utility of this framework in two respects. First, we give the only proof of which we are aware that a class of deep neural networks can be trained using gradient descent to global optima even when such optima only exist at infinity, as is the case for the cross-entropy cost. Second, we identify a novel causal mechanism by which skip connections accelerate training, which we verify predictively with ResNets on MNIST, CIFAR10, CIFAR100 and ImageNet.

Deep Learning/Theory

Jingfeng Wu, Vladimir Braverman, Jason Lee

Recent research has observed that in machine learning optimization, gradient descent (GD) often operates at the edge of stability (EoS) [Cohen et al., 2021], where the stepsizes are set to be large, resulting in non-monotonic losses induced by the GD iterates. This paper studies the convergence and implicit bias of constant-stepsize GD for logistic regression on linearly separable data in the EoS regime. Despite the presence of local oscillations, we prove that the logistic loss can be minimized by GD with any constant stepsize over a long time scale. Furthermore, we prove that with any constant stepsize, the GD iterates tend to infinity when projected to a max-margin direction (the hard-margin SVM direction) and converge to a fixed vector that minimizes a strongly convex potential when projected to the orthogonal complement of the max-margin direction. In contrast, we also show that in the EoS regime, GD iterates may diverge catastrophically under the exponential loss, highlighting the superiority of the logistic loss. These theoretical findings are in line with numerical simulations and complement existing theories on the convergence and implicit bias of GD for logistic regression, which are only applicable when the stepsizes are sufficiently small.

Deep Learning/Theory

Aditya Vardhan Varre, Maria-Luiza Vladarean, Loucas PILLAUD-VIVIEN, Nicolas Flammarion

This paper studies the behaviour of two-layer fully connected networks with linear activations trained with gradient flow on the square loss. We show how the optimization process carries an implicit bias on the parameters that depends on the scale of its initialization. The main result of the paper is a variational characterization of the loss minimizers retrieved by the gradient flow for a specific initialization shape. This characterization reveals that, in the small scale initialization regime, the linear neural network's hidden layer is biased toward having a low-rank structure. To complement our results, we showcase a hidden mirror flow that tracks the dynamics of the singular values of the weights matrices and describe their time evolution. We support our findings with numerical experiments illustrating the phenomena.

Deep Learning/Theory

Dian Wang, Xupeng Zhu, Jung Yeon Park, Mingxi Jia, Guanang Su, Robert Platt, Robin Walters

Although equivariant machine learning has proven effective at many tasks, success depends heavily on the assumption that the ground truth function is symmetric over the entire domain matching the symmetry in an equivariant neural network. A missing piece in the equivariant learning literature is the analysis of equivariant networks when symmetry exists only partially in the domain. In this work, we present a general theory for such a situation. We propose pointwise definitions of correct, incorrect, and extrinsic equivariance, which allow us to quantify continuously the degree of each type of equivariance a function displays. We then study the impact of various degrees of incorrect or extrinsic symmetry on model error. We prove error lower bounds for invariant or equivariant networks in classification or regression settings with partially incorrect symmetry. We also analyze the potentially harmful effects of extrinsic equivariance. Experiments validate these results in three different environments.

Deep Learning/Theory

Bochen Lyu, Zhanxing Zhu

Studying the implicit bias of gradient descent (GD) and stochastic gradient descent (SGD) is critical to unveil the underlying mechanism of deep learning. Unfortunately, even for standard linear networks in regression setting, a comprehensive characterization of the implicit bias is still an open problem. This paper proposes to investigate a new proxy model of standard linear network, rank-1 linear network, where each weight matrix is parameterized as a rank-1 form. For over-parameterized regression problem, we precisely analyze the implicit bias of GD and SGD---by identifying a “potential” function such that GD converges to its minimizer constrained by zero training error (i.e., interpolation solution), and further characterizing the role of the noise introduced by SGD in perturbing the form of this potential. Our results explicitly connect the depth of the network and the initialization with the implicit bias of GD and SGD. Furthermore, we emphasize a new implicit bias of SGD jointly induced by stochasticity and over-parameterization, which can reduce the dependence of the SGD's solution on the initialization. Our findings regarding the implicit bias are different from that of a recently popular model, the diagonal linear network. We highlight that the induced bias of our rank-1 model is more consistent with standard linear network while the diagonal one is not. This suggests that the proposed rank-1 linear network might be a plausible proxy for standard linear net.

Deep Learning/Theory

Jiancong Xiao, Ruoyu Sun, Zhi-Quan Luo

Deep neural networks (DNNs) are vulnerable to adversarial attacks. It is found empirically that adversarially robust generalization is crucial in establishing defense algorithms against adversarial attacks. Therefore, it is interesting to study the theoretical guarantee of robust generalization. This paper focuses on norm-based complexity, based on a PAC-Bayes approach (Neyshabur et al., 2017). The main challenge lies in extending the key ingredient, which is a weight perturbation bound in standard settings, to the robust settings. Existing attempts heavily rely on additional strong assumptions, leading to loose bounds. In this paper, we address this issue and provide a spectrally-normalized robust generalization bound for DNNs. Compared to existing bounds, our bound offers two significant advantages: Firstly, it does not depend on additional assumptions. Secondly, it is considerably tighter, aligning with the bounds of standard generalization. Therefore, our result provides a different perspective on understanding robust generalization: The mismatch terms between standard and robust generalization bounds shown in previous studies do not contribute to the poor robust generalization. Instead, these disparities solely due to mathematical issues. Finally, we extend the main result to adversarial robustness against general non-$\ell_p$ attacks and other neural network architectures.

Deep Learning/Theory

Ajil Jalal, Justin Kang, Ananya Uppal, Kannan Ramchandran, Eric Price

A conditional generative model is a method for sampling from a conditional distribution $p(y \mid x)$. For example, one may want to sample an image of a cat given the label ``cat''. A feed-forward conditional generative model is a function $g(x, z)$ that takes the input $x$ and a random seed $z$, and outputs a sample $y$ from $p(y \mid x)$. Ideally the distribution of outputs $(x, g(x, z))$ would be close in total variation to the ideal distribution $(x, y)$.Generalization bounds for other learning models require assumptions on the distribution of $x$, even in simple settings like linear regression with Gaussian noise. We show these assumptions are unnecessary in our model, for both linear regression and single-layer ReLU networks. Given samples $(x, y)$, we show how to learn a 1-layer ReLU conditional generative model in total variation. As our result has no assumption on the distribution of inputs $x$, if we are given access to the internal activations of a deep generative model, we can compose our 1-layer guarantee to progressively learn the deep model using a near-linear number of samples.

Deep Learning/Theory

Rajat Vadiraj Dwaraknath, Tolga Ergen, Mert Pilanci

Recently, theoretical analyses of deep neural networks have broadly focused on two directions: 1) Providing insight into neural network training by SGD in the limit of infinite hidden-layer width and infinitesimally small learning rate (also known as gradient flow) via the Neural Tangent Kernel (NTK), and 2) Globally optimizing the regularized training objective via cone-constrained convex reformulations of ReLU networks. The latter research direction also yielded an alternative formulation of the ReLU network, called a gated ReLU network, that is globally optimizable via efficient unconstrained convex programs. In this work, we interpret the convex program for this gated ReLU network as a Multiple Kernel Learning (MKL) model with a weighted data masking feature map and establish a connection to the NTK. Specifically, we show that for a particular choice of mask weights that do not depend on the learning targets, this kernel is equivalent to the NTK of the gated ReLU network on the training data. A consequence of this lack of dependence on the targets is that the NTK cannot perform better than the optimal MKL kernel on the training set. By using iterative reweighting, we improve the weights induced by the NTK to obtain the optimal MKL kernel which is equivalent to the solution of the exact convex reformulation of the gated ReLU network. We also provide several numerical simulations corroborating our theory. Additionally, we provide an analysis of the prediction error of the resulting optimal kernel via consistency results for the group lasso.

Deep Learning/Theory

Jiyoung Park, Ian Pelakh, Stephan Wojtowytsch

We investigate how shallow ReLU networks interpolate between known regions. Our analysis shows that empirical risk minimizers converge to a minimum norm interpolant as the number of data points and parameters tends to infinity when a weight decay regularizer is penalized with a coefficient which vanishes at a precise rate as the network width and the number of data points grow. With and without explicit regularization, we numerically study the implicit bias of common optimization algorithms towards known minimum norm interpolants.

Deep Learning/Theory

Hengyu Fu, Tianyu Guo, Yu Bai, Song Mei

Attention layers---which map a sequence of inputs to a sequence of outputs---are core building blocks of the Transformer architecture which has achieved significant breakthroughs in modern artificial intelligence. This paper presents a rigorous theoretical study on the learning and generalization of a single multi-head attention layer, with a sequence of key vectors and a separate query vector as input. We consider the random feature setting where the attention layer has a large number of heads, with randomly sampled frozen query and key matrices, and trainable value matrices. We show that such a random-feature attention layer can express a broad class of target functions that are permutation invariant to the key vectors. We further provide quantitative excess risk bounds for learning these target functions from finite samples, using random feature attention with finitely many heads.Our results feature several implications unique to the attention structure compared with existing random features theory for neural networks, such as (1) Advantages in the sample complexity over standard two-layer random-feature networks; (2) Concrete and natural classes of functions that can be learned efficiently by a random-feature attention layer; and (3) The effect of the sampling distribution of the query-key weight matrix (the product of the query and key matrix), where Gaussian random weights with a non-zero mean result in better sample complexities over the zero-mean counterpart for learning certain natural target functions. Experiments on simulated data corroborate our theoretical findings and further illustrate the interplay between the sample size and the complexity of the target function.

Deep Learning/Theory

Amir Joudaki, Hadi Daneshmand, Francis Bach

In this paper, we explore the structure of the penultimate Gram matrix in deep neural networks, which contains the pairwise inner products of outputs corresponding to a batch of inputs. In several architectures it has been observed that this Gram matrix becomes degenerate with depth at initialization, which dramatically slows training. Normalization layers, such as batch or layer normalization, play a pivotal role in preventing the rank collapse issue. Despite promising advances, the existing theoretical results do not extend to layer normalization, which is widely used in transformers, and can not quantitatively characterize the role of non-linear activations. To bridge this gap, we prove that layer normalization, in conjunction with activation layers, biases the Gram matrix of a multilayer perceptron towards the identity matrix at an exponential rate with depth at initialization. We quantify this rate using the Hermite expansion of the activation function.

Deep Learning/Theory

Yan Dai, Kwangjun Ahn, Suvrit Sra

Sharpness-Aware Minimization (SAM) is a recently proposed gradient-based optimizer (Foret et al., ICLR 2021) that greatly improves the prediction performance of deep neural networks. Consequently, there has been a surge of interest in explaining its empirical success. We focus, in particular, on understanding ***the role played by normalization***, a key component of the SAM updates. We theoretically and empirically study the effect of normalization in SAM for both convex and non-convex functions, revealing two key roles played by normalization: i) it helps in stabilizing the algorithm; and ii) it enables the algorithm to drift along a continuum (manifold) of minima -- a property identified by recent theoretical works that is the key to better performance. We further argue that these two properties of normalization make SAM robust against the choice of hyper-parameters, supporting the practicality of SAM. Our conclusions are backed by various experiments.

Deep Learning/Theory

Eshaan Nichani, Alex Damian, Jason Lee

One of the central questions in the theory of deep learning is to understand how neural networks learn hierarchical features. The ability of deep networks to extract salient features is crucial to both their outstanding generalization ability and the modern deep learning paradigm of pretraining and finetuneing. However, this feature learning process remains poorly understood from a theoretical perspective, with existing analyses largely restricted to two-layer networks. In this work we show that three-layer neural networks have provably richer feature learning capabilities than two-layer networks. We analyze the features learned by a three-layer network trained with layer-wise gradient descent, and present a general purpose theorem which upper bounds the sample complexity and width needed to achieve low test error when the target has specific hierarchical structure. We instantiate our framework in specific statistical learning settings -- single-index models and functions of quadratic features -- and show that in the latter setting three-layer networks obtain a sample complexity improvement over all existing guarantees for two-layer networks. Crucially, this sample complexity improvement relies on the ability of three-layer networks to efficiently learn *nonlinear* features. We then establish a concrete optimization-based depth separation by constructing a function which is efficiently learnable via gradient descent on a three-layer network, yet cannot be learned efficiently by a two-layer network. Our work makes progress towards understanding the provable benefit of three-layer neural networks over two-layer networks in the feature learning regime.

Deep Learning/Theory

Behrooz Tahmasebi, Stefanie Jegelka

In practice, encoding invariances into models improves sample complexity. In this work, we study this phenomenon from a theoretical perspective. In particular, we provide minimax optimal rates for kernel ridge regression on compact manifolds, with a target function that is invariant to a group action on the manifold. Our results hold for any smooth compact Lie group action, even groups of positive dimension. For a finite group, the gain effectively multiplies the number of samples by the group size. For groups of positive dimension, the gain is observed by a reduction in the manifold's dimension, in addition to a factor proportional to the volume of the quotient space. Our proof takes the viewpoint of differential geometry, in contrast to the more common strategy of using invariant polynomials. This new geometric viewpoint on learning with invariances may be of independent interest.

Deep Learning/Theory

Alex Damian, Eshaan Nichani, Rong Ge, Jason Lee

We focus on the task of learning a single index model $\sigma(w^\star \cdot x)$ with respect to the isotropic Gaussian distribution in $d$ dimensions. Prior work has shown that the sample complexity of learning $w^\star$ is governed by the information exponent $k^\star$ of the link function $\sigma$, which is defined as the index of the first nonzero Hermite coefficient of $\sigma$. Ben Arous et al. (2021) showed that $n \gtrsim d^{k^\star-1}$ samples suffice for learning $w^\star$ and that this is tight for online SGD. However, the CSQ lower bound for gradient based methods only shows that $n \gtrsim d^{k^\star/2}$ samples are necessary. In this work, we close the gap between the upper and lower bounds by showing that online SGD on a smoothed loss learns $w^\star$ with $n \gtrsim d^{k^\star/2}$ samples. We also draw connections to statistical analyses of tensor PCA and to the implicit regularization effects of minibatch SGD on empirical losses.

Deep Learning/Theory

Milad Sefidgaran, Abdellatif Zaidi, Piotr Krasnowski

A major challenge in designing efficient statistical supervised learning algorithms is finding representations that perform well not only on available training samples but also on unseen data. While the study of representation learning has spurred much interest, most existing such approaches are heuristic; and very little is known about theoretical generalization guarantees. For example, the information bottleneck method seeks a good generalization by finding a minimal description of the input that is maximally informative about the label variable, where minimality and informativeness are both measured by Shannon’s mutual information. In this paper, we establish a compressibility framework that allows us to derive upper bounds on the generalization error of a representation learning algorithm in terms of the ``Minimum Description Length'' (MDL) of the labels or the latent variables (representations). Rather than the mutual information between the encoder’s input and the representation, which is often believed to reflect the algorithm’s generalization capability in the related literature but in fact, falls short of doing so, our new bounds involve the "multi-letter" relative entropy between the distribution of the representations (or labels) of the training and test sets and a fixed prior. In particular, these new bounds reflect the structure of the encoder and are not vacuous for deterministic algorithms. Our compressibility approach, which is information-theoretic in nature, builds upon that of Blum-Langford for PAC-MDL bounds and introduces two essential ingredients: block-coding and lossy-compression. The latter allows our approach to subsume the so-called geometrical compressibility as a special case. To the best knowledge of the authors, the established generalization bounds are the first of their kind for Information Bottleneck type encoders and representation learning. Finally, we partly exploit the theoretical results by introducing a new data-dependent prior. Numerical simulations illustrate the advantages of well-chosen such priors over classical priors used in IB.

Deep Learning/Theory

Daniel Bertschinger, Christoph Hertrich, Paul Jungeblut, Tillmann Miltzow, Simon Weber

We consider the algorithmic problem of finding the optimal weights and biases for a two-layer fully connected neural network to fit a given set of data points, also known as empirical risk minimization. We show that the problem is $\exists\mathbb{R}$-complete. This complexity class can be defined as the set of algorithmic problems that are polynomial-time equivalent to finding real roots of a multivariate polynomial with integer coefficients. Furthermore, we show that arbitrary algebraic numbers are required as weights to be able to train some instances to optimality, even if all data points are rational. Our result already applies to fully connected instances with two inputs, two outputs, and one hidden layer of ReLU neurons. Thereby, we strengthen a result by Abrahamsen, Kleist and Miltzow [NeurIPS 2021]. A consequence of this is that a combinatorial search algorithm like the one by Arora, Basu, Mianjy and Mukherjee [ICLR 2018] is impossible for networks with more than one output dimension, unless $\text{NP} = \exists\mathbb{R}$.

Deep Learning/Theory

Chen Zeno, Greg Ongie, Yaniv Blumenfeld, Nir Weinberger, Daniel Soudry

Neural network (NN) denoisers are an essential building block in many common tasks, ranging from image reconstruction to image generation. However, the success of these models is not well understood from a theoretical perspective. In this paper, we aim to characterize the functions realized by shallow ReLU NN denoisers --- in the common theoretical setting of interpolation (i.e., zero training loss) with a minimal representation cost (i.e., minimal $\ell^2$ norm weights). First, for univariate data, we derive a closed form for the NN denoiser function, find it is contractive toward the clean data points, and prove it generalizes better than the empirical MMSE estimator at a low noise level. Next, for multivariate data, we find the NN denoiser functions in a closed form under various geometric assumptions on the training data: data contained in a low-dimensional subspace, data contained in a union of one-sided rays, or several types of simplexes. These functions decompose into a sum of simple rank-one piecewise linear interpolations aligned with edges and/or faces connecting training samples. We empirically verify this alignment phenomenon on synthetic data and real images.

Deep Learning/Theory

Paul Geuchen, Felix Voigtlaender

Complex-valued neural networks (CVNNs) have recently shown promising empirical success, for instance for increasing the stability of recurrent neural networks and for improving the performance in tasks with complex-valued inputs, such as MRI fingerprinting. While the overwhelming success of Deep Learning in the real-valued case is supported by a growing mathematical foundation, such a foundation is still largely lacking in the complex-valued case. We thus analyze the expressivity of CVNNs by studying their approximation properties. Our results yield the first quantitative approximation bounds for CVNNs that apply to a wide class of activation functions including the popular modReLU and complex cardioid activation functions. Precisely, our results apply to any activation function that is smooth but not polyharmonic on some non-empty open set; this is the natural generalization of the class of smooth and non-polynomial activation functions to the complex setting. Our main result shows that the approximation error scales as $m^{-k/(2n)}$ for $m \to \infty$ where $m$ is the number of neurons, $k$ the smoothness of the target function and $n$ is the (complex) input dimension. Under a natural continuity assumption, we show that this rate is optimal; we further discuss the optimality when dropping this assumption. Moreover, we prove that the problem of approximating $C^k$-functions using continuous approximation methods unavoidably suffers from the curse of dimensionality.