Skip to yearly menu bar Skip to main content


Oral

Smoothing the Landscape Boosts the Signal for SGD: Optimal Sample Complexity for Learning Single Index Models

Alex Damian · Eshaan Nichani · Rong Ge · Jason Lee

Hall C2 (level 1 gate 9 south of food court)

Abstract: We focus on the task of learning a single index model σ(wx)σ(wx) with respect to the isotropic Gaussian distribution in dd dimensions. Prior work has shown that the sample complexity of learning ww is governed by the information exponent kk of the link function σσ, which is defined as the index of the first nonzero Hermite coefficient of σσ. Ben Arous et al. (2021) showed that ndk1 samples suffice for learning w and that this is tight for online SGD. However, the CSQ lower bound for gradient based methods only shows that ndk/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 with ndk/2 samples. We also draw connections to statistical analyses of tensor PCA and to the implicit regularization effects of minibatch SGD on empirical losses.

Chat is not available.