Learning a Single Neuron with Adversarial Label Noise via Gradient Descent

TL;DR

This paper develops efficient algorithms for learning a single neuron with monotone activation functions under adversarial label noise, achieving near-optimal approximation guarantees for log-concave distributions using gradient descent.

Contribution

It introduces the first polynomial-time constant-factor approximation algorithms for single neuron learning with adversarial noise under broad distribution classes, including log-concave distributions.

Findings

First polynomial-time constant-factor approximation for logistic activation.

Sample complexity matches known lower bounds up to polylog factors.

Algorithms are simple gradient descent methods with novel structural analysis.

Abstract

We study the fundamental problem of learning a single neuron, i.e., a function of the form $\mathbf{x}\mapsto\sigma(\mathbf{w}\cdot\mathbf{x})$ for monotone activations $\sigma:\mathbb{R}\mapsto\mathbb{R}$, with respect to the $L_2^2$-loss in the presence of adversarial label noise. Specifically, we are given labeled examples from a distribution $D$ on $(\mathbf{x}, y)\in\mathbb{R}^d \times \mathbb{R}$ such that there exists $\mathbf{w}^\ast\in\mathbb{R}^d$ achieving $F(\mathbf{w}^\ast)=\epsilon$, where $F(\mathbf{w})=\mathbf{E}_{(\mathbf{x},y)\sim D}[(\sigma(\mathbf{w}\cdot \mathbf{x})-y)^2]$. The goal of the learner is to output a hypothesis vector $\mathbf{w}$ such that $F(\mathbb{w})=C\, \epsilon$ with high probability, where $C>1$ is a universal constant. As our main contribution, we give efficient constant-factor approximate learners for a broad class of distributions (including log-concave distributions) and activation functions. Concretely, for the class of isotropic log-concave distributions, we obtain the following important corollaries: For the logistic activation, we obtain the first polynomial-time constant factor approximation (even under the Gaussian distribution). Our algorithm has sample complexity $\widetilde{O}(d/\epsilon)$, which is tight within polylogarithmic factors. For the ReLU activation, we give an efficient algorithm with sample complexity $\tilde{O}(d\, \polylog(1/\epsilon))$. Prior to our work, the best known constant-factor approximate learner had sample complexity $\tilde{\Omega}(d/\epsilon)$. In both of these settings, our algorithms are simple, performing gradient-descent on the (regularized) $L_2^2$-loss. The correctness of our algorithms relies on novel structural results that we establish, showing that (essentially all) stationary points of the underlying non-convex loss are approximately optimal.