NeurIPS2022

Learning (Very) Simple Generative Models Is Hard

Sitan Chen, Jerry Li, Yuanzhi Li

12 citations

Abstract

Motivated by the recent empirical successes of deep generative models, we study the computational complexity of the following unsupervised learning problem. For an unknown neural network F:RdRdF:\mathbb{R}^d\to\mathbb{R}^{d'}, let DD be the distribution over Rd\mathbb{R}^{d'} given by pushing the standard Gaussian N(0,Idd)\mathcal{N}(0,\textrm{Id}_d) through FF. Given i.i.d. samples from DD, the goal is to output any distribution close to DD in statistical distance. We show under the statistical query (SQ) model that no polynomial-time algorithm can solve this problem even when the output coordinates of FF are one-hidden-layer ReLU networks with log(d)\log(d) neurons. Previously, the best lower bounds for this problem simply followed from lower bounds for supervised learning and required at least two hidden layers and poly(d)\mathrm{poly}(d) neurons [Daniely-Vardi '21, Chen-Gollakota-Klivans-Meka '22]. The key ingredient in our proof is an ODE-based construction of a compactly supported, piecewise-linear function ff with polynomially-bounded slopes such that the pushforward of N(0,1)\mathcal{N}(0,1) under ff matches all low-degree moments of N(0,1)\mathcal{N}(0,1).