arXiv subjects

Zeyuan Allen-Zhu

Publications and source records attributed to Zeyuan Allen-Zhu.

At least 37 records · Page 2Linked to original sources

arXiv2020

What Can ResNet Learn Efficiently, Going Beyond Kernels?

How can neural networks such as ResNet efficiently learn CIFAR-10 with test accuracy more than 96%, while other methods, especially kernel methods, fall relatively behind? Can we more provide theoretical justifications for this gap? Recently, there is an influential line of work relating neural networks to kernels in the over-parameterized regime, proving they can learn certain concept class that is also learnable by kernels with similar test error. Yet, can neural networks provably learn some concept class BETTER than kernels? We answer this positively in the distribution-free setting. We prove neural networks can efficiently learn a notable class of functions, including those defined by three-layer residual networks with smooth activations, without any distributional assumption. At the same time, we prove there are simple functions in this class such that with the same number of training examples, the test error obtained by neural networks can be MUCH SMALLER than ANY kernel method, including neural tangent kernels (NTK). The main intuition is that multi-layer neural networks can implicitly perform hierarchical learning using different layers, which reduces the sample complexity comparing to "one-shot" learning algorithms such as kernel methods. In a follow-up work [2], this theory of hierarchical learning is further strengthened to incorporate the "backward feature correction" process when training deep networks. In the end, we also prove a computation complexity advantage of ResNet with respect to other learning methods including linear regression over arbitrary feature mappings.

cs.LG↗

arXiv2019

A Convergence Theory for Deep Learning via Over-Parameterization

Deep neural networks (DNNs) have demonstrated dominating performance in many fields; since AlexNet, networks used in practice are going wider and deeper. On the theoretical side, a long line of works has been focusing on training neural networks with one hidden layer. The theory of multi-layer networks remains largely unsettled. In this work, we prove why stochastic gradient descent (SGD) can find $\textit{global minima}$ on the training objective of DNNs in $\textit{polynomial time}$. We only make two assumptions: the inputs are non-degenerate and the network is over-parameterized. The latter means the network width is sufficiently large: $\textit{polynomial}$ in $L$, the number of layers and in $n$, the number of samples. Our key technique is to derive that, in a sufficiently large neighborhood of the random initialization, the optimization landscape is almost-convex and semi-smooth even with ReLU activations. This implies an equivalence between over-parameterized neural networks and neural tangent kernel (NTK) in the finite (and polynomial) width setting. As concrete examples, starting from randomly initialized weights, we prove that SGD can attain 100% training accuracy in classification tasks, or minimize regression loss in linear convergence speed, with running time polynomial in $n,L$. Our theory applies to the widely-used but non-smooth ReLU activation, and to any smooth and possibly non-convex loss functions. In terms of network architectures, our theory at least applies to fully-connected neural networks, convolutional neural networks (CNN), and residual neural networks (ResNet).

cs.LG↗

arXiv2019

The Lingering of Gradients: Theory and Applications

Classically, the time complexity of a first-order method is estimated by its number of gradient computations. In this paper, we study a more refined complexity by taking into account the `lingering' of gradients: once a gradient is computed at $x_k$, the additional time to compute gradients at $x_{k+1},x_{k+2},\dots$ may be reduced. We show how this improves the running time of several first-order methods. For instance, if the `additional time' scales linearly with respect to the traveled distance, then the `convergence rate' of gradient descent can be improved from $1/T$ to $\exp(-T^{1/3})$. On the application side, we solve a hypothetical revenue management problem on the Yahoo! Front Page Today Module with 4.6m users to $10^{-6}$ error using only 6 passes of the dataset; and solve a real-life support vector machine problem to an accuracy that is two orders of magnitude better comparing to the state-of-the-art algorithm.

math.OC↗

arXiv2019

Can SGD Learn Recurrent Neural Networks with Provable Generalization?

Recurrent Neural Networks (RNNs) are among the most popular models in sequential data analysis. Yet, in the foundational PAC learning language, what concept class can it learn? Moreover, how can the same recurrent unit simultaneously learn functions from different input tokens to different output tokens, without affecting each other? Existing generalization bounds for RNN scale exponentially with the input length, significantly limiting their practical implications. In this paper, we show using the vanilla stochastic gradient descent (SGD), RNN can actually learn some notable concept class efficiently, meaning that both time and sample complexity scale polynomially in the input length (or almost polynomially, depending on the concept). This concept class at least includes functions where each output token is generated from inputs of earlier tokens using a smooth two-layer neural network.

cs.LG↗

arXiv2019

On the Convergence Rate of Training Recurrent Neural Networks

How can local-search methods such as stochastic gradient descent (SGD) avoid bad local minima in training multi-layer neural networks? Why can they fit random labels even given non-convex and non-smooth architectures? Most existing theory only covers networks with one hidden layer, so can we go deeper? In this paper, we focus on recurrent neural networks (RNNs) which are multi-layer networks widely used in natural language processing. They are harder to analyze than feedforward neural networks, because the $\textit{same}$ recurrent unit is repeatedly applied across the entire time horizon of length $L$, which is analogous to feedforward networks of depth $L$. We show when the number of neurons is sufficiently large, meaning polynomial in the training data size and in $L$, then SGD is capable of minimizing the regression loss in the linear convergence rate. This gives theoretical evidence of how RNNs can memorize data. More importantly, in this paper we build general toolkits to analyze multi-layer networks with ReLU activations. For instance, we prove why ReLU activations can prevent exponential gradient explosion or vanishing, and build a perturbation theory to analyze first-order approximation of multi-layer networks.

cs.LG↗

arXiv2018

Natasha: Faster Non-Convex Stochastic Optimization Via Strongly Non-Convex Parameter

Given a nonconvex function that is an average of $n$ smooth functions, we design stochastic first-order methods to find its approximate stationary points. The convergence of our new methods depends on the smallest (negative) eigenvalue $-σ$ of the Hessian, a parameter that describes how nonconvex the function is. Our methods outperform known results for a range of parameter $σ$, and can be used to find approximate local minima. Our result implies an interesting dichotomy: there exists a threshold $σ_0$ so that the currently fastest methods for $σ>σ_0$ and for $σ<σ_0$ have different behaviors: the former scales with $n^{2/3}$ and the latter scales with $n^{3/4}$.

math.OC↗

arXiv2018

Katyusha: The First Direct Acceleration of Stochastic Gradient Methods

Nesterov's momentum trick is famously known for accelerating gradient descent, and has been proven useful in building fast iterative algorithms. However, in the stochastic setting, counterexamples exist and prevent Nesterov's momentum from providing similar acceleration, even if the underlying problem is convex and finite-sum. We introduce $\mathtt{Katyusha}$, a direct, primal-only stochastic gradient method to fix this issue. In convex finite-sum stochastic optimization, $\mathtt{Katyusha}$ has an optimal accelerated convergence rate, and enjoys an optimal parallel linear speedup in the mini-batch setting. The main ingredient is $\textit{Katyusha momentum}$, a novel "negative momentum" on top of Nesterov's momentum. It can be incorporated into a variance-reduction based algorithm and speed it up, both in terms of $\textit{sequential and parallel}$ performance. Since variance reduction has been successfully applied to a growing list of practical problems, our paper suggests that in each of such cases, one could potentially try to give Katyusha a hug.

math.OC↗

arXiv2018

Is Q-learning Provably Efficient?

Model-free reinforcement learning (RL) algorithms, such as Q-learning, directly parameterize and update value functions or policies without explicitly modeling the environment. They are typically simpler, more flexible to use, and thus more prevalent in modern deep RL than model-based approaches. However, empirical work has suggested that model-free algorithms may require more samples to learn [Deisenroth and Rasmussen 2011, Schulman et al. 2015]. The theoretical question of "whether model-free algorithms can be made sample efficient" is one of the most fundamental questions in RL, and remains unsolved even in the basic scenario with finitely many states and actions. We prove that, in an episodic MDP setting, Q-learning with UCB exploration achieves regret $\tilde{O}(\sqrt{H^3 SAT})$, where $S$ and $A$ are the numbers of states and actions, $H$ is the number of steps per episode, and $T$ is the total number of steps. This sample efficiency matches the optimal regret that can be achieved by any model-based approach, up to a single $\sqrt{H}$ factor. To the best of our knowledge, this is the first analysis in the model-free setting that establishes $\sqrt{T}$ regret without requiring access to a "simulator."

cs.LG↗

arXiv2018

Natasha 2: Faster Non-Convex Optimization Than SGD

We design a stochastic algorithm to train any smooth neural network to $\varepsilon$-approximate local minima, using $O(\varepsilon^{-3.25})$ backpropagations. The best result was essentially $O(\varepsilon^{-4})$ by SGD. More broadly, it finds $\varepsilon$-approximate local minima of any smooth nonconvex function in rate $O(\varepsilon^{-3.25})$, with only oracle access to stochastic gradients.

math.OC↗

arXiv2018

Neon2: Finding Local Minima via First-Order Oracles

We propose a reduction for non-convex optimization that can (1) turn an stationary-point finding algorithm into an local-minimum finding one, and (2) replace the Hessian-vector product computations with only gradient computations. It works both in the stochastic and the deterministic settings, without hurting the algorithm's performance. As applications, our reduction turns Natasha2 into a first-order method without hurting its performance. It also converts SGD, GD, SCSG, and SVRG into algorithms finding approximate local minima, outperforming some best known results.

cs.LG↗

arXiv2018

Operator Scaling via Geodesically Convex Optimization, Invariant Theory and Polynomial Identity Testing

We propose a new second-order method for geodesically convex optimization on the natural hyperbolic metric over positive definite matrices. We apply it to solve the operator scaling problem in time polynomial in the input size and logarithmic in the error. This is an exponential improvement over previous algorithms which were analyzed in the usual Euclidean, "commutative" metric (for which the above problem is not convex). Our method is general and applicable to other settings. As a consequence, we solve the equivalence problem for the left-right group action underlying the operator scaling problem. This yields a deterministic polynomial-time algorithm for a new class of Polynomial Identity Testing (PIT) problems, which was the original motivation for studying operator scaling.

cs.DS↗

arXiv2018

Byzantine Stochastic Gradient Descent

This paper studies the problem of distributed stochastic optimization in an adversarial setting where, out of the $m$ machines which allegedly compute stochastic gradients every iteration, an $α$-fraction are Byzantine, and can behave arbitrarily and adversarially. Our main result is a variant of stochastic gradient descent (SGD) which finds $\varepsilon$-approximate minimizers of convex functions in $T = \tilde{O}\big( \frac{1}{\varepsilon^2 m} + \frac{α^2}{\varepsilon^2} \big)$ iterations. In contrast, traditional mini-batch SGD needs $T = O\big( \frac{1}{\varepsilon^2 m} \big)$ iterations, but cannot tolerate Byzantine failures. Further, we provide a lower bound showing that, up to logarithmic factors, our algorithm is information-theoretically optimal both in terms of sampling complexity and time complexity.

cs.LG↗

arXiv2018

Nearly Linear-Time Packing and Covering LP Solvers

Packing and covering linear programs (PC-LPs) form an important class of linear programs (LPs) across computer science, operations research, and optimization. In 1993, Luby and Nisan constructed an iterative algorithm for approximately solving PC-LPs in nearly linear time, where the time complexity scales nearly linearly in $N$, the number of nonzero entries of the matrix, and polynomially in $\varepsilon$, the (multiplicative) approximation error. Unfortunately, all existing nearly linear-time algorithms for solving PC-LPs require time at least proportional to $\varepsilon^{-2}$. In this paper, we break this longstanding barrier by designing a packing solver that runs in time $\tilde{O}(N \varepsilon^{-1})$ and covering LP solver that runs in time $\tilde{O}(N \varepsilon^{-1.5})$. Our packing solver can be extended to run in time $\tilde{O}(N \varepsilon^{-1})$ for a class of well-behaved covering programs. In a follow-up work, Wang et al. showed that all covering LPs can be converted into well-behaved ones by a reduction that blows up the problem size only logarithmically. At high level, these two algorithms can be described as linear couplings of several first-order descent steps. This is an application of our linear coupling technique to problems that are not amenable to blackbox applications known iterative algorithms in convex optimization.

cs.DS↗

arXiv2018

Katyusha X: Practical Momentum Method for Stochastic Sum-of-Nonconvex Optimization

The problem of minimizing sum-of-nonconvex functions (i.e., convex functions that are average of non-convex ones) is becoming increasingly important in machine learning, and is the core machinery for PCA, SVD, regularized Newton's method, accelerated non-convex optimization, and more. We show how to provably obtain an accelerated stochastic algorithm for minimizing sum-of-nonconvex functions, by $\textit{adding one additional line}$ to the well-known SVRG method. This line corresponds to momentum, and shows how to directly apply momentum to the finite-sum stochastic minimization of sum-of-nonconvex functions. As a side result, our method enjoys linear parallel speed-up using mini-batch.

cs.LG↗

arXiv2018

Make the Minority Great Again: First-Order Regret Bound for Contextual Bandits

Regret bounds in online learning compare the player's performance to $L^*$, the optimal performance in hindsight with a fixed strategy. Typically such bounds scale with the square root of the time horizon $T$. The more refined concept of first-order regret bound replaces this with a scaling $\sqrt{L^*}$, which may be much smaller than $\sqrt{T}$. It is well known that minor variants of standard algorithms satisfy first-order regret bounds in the full information and multi-armed bandit settings. In a COLT 2017 open problem, Agarwal, Krishnamurthy, Langford, Luo, and Schapire raised the issue that existing techniques do not seem sufficient to obtain first-order regret bounds for the contextual bandit problem. In the present paper, we resolve this open problem by presenting a new strategy based on augmenting the policy space.

cs.LG↗

arXiv2017

Near-Optimal Discrete Optimization for Experimental Design: A Regret Minimization Approach

The experimental design problem concerns the selection of k points from a potentially large design pool of p-dimensional vectors, so as to maximize the statistical efficiency regressed on the selected k design points. Statistical efficiency is measured by optimality criteria, including A(verage), D(eterminant), T(race), E(igen), V(ariance) and G-optimality. Except for the T-optimality, exact optimization is NP-hard. We propose a polynomial-time regret minimization framework to achieve a $(1+\varepsilon)$ approximation with only $O(p/\varepsilon^2)$ design points, for all the optimality criteria above. In contrast, to the best of our knowledge, before our work, no polynomial-time algorithm achieves $(1+\varepsilon)$ approximations for D/E/G-optimality, and the best poly-time algorithm achieving $(1+\varepsilon)$-approximation for A/V-optimality requires $k = Ω(p^2/\varepsilon)$ design points.

stat.ML↗

arXiv2017

Linear Convergence of a Frank-Wolfe Type Algorithm over Trace-Norm Balls

We propose a rank-$k$ variant of the classical Frank-Wolfe algorithm to solve convex optimization over a trace-norm ball. Our algorithm replaces the top singular-vector computation ($1$-SVD) in Frank-Wolfe with a top-$k$ singular-vector computation ($k$-SVD), which can be done by repeatedly applying $1$-SVD $k$ times. Alternatively, our algorithm can be viewed as a rank-$k$ restricted version of projected gradient descent. We show that our algorithm has a linear convergence rate when the objective function is smooth and strongly convex, and the optimal solution has rank at most $k$. This improves the convergence rate and the total time complexity of the Frank-Wolfe method and its variants.

cs.LG↗

arXiv2017

Follow the Compressed Leader: Faster Online Learning of Eigenvectors and Faster MMWU

The online problem of computing the top eigenvector is fundamental to machine learning. In both adversarial and stochastic settings, previous results (such as matrix multiplicative weight update, follow the regularized leader, follow the compressed leader, block power method) either achieve optimal regret but run slow, or run fast at the expense of loosing a $\sqrt{d}$ factor in total regret where $d$ is the matrix dimension. We propose a $\textit{follow-the-compressed-leader (FTCL)}$ framework which achieves optimal regret without sacrificing the running time. Our idea is to "compress" the matrix strategy to dimension 3 in the adversarial setting, or dimension 1 in the stochastic setting. These respectively resolve two open questions regarding the design of optimal and efficient algorithms for the online eigenvector problem.

cs.LG↗