Deep Double Descent: Where Bigger Models and More Data Hurt

TL;DR

This paper investigates the double descent phenomenon in deep learning, revealing that larger models and more data can sometimes degrade performance, and introduces a new complexity measure to unify these observations.

Contribution

It introduces the concept of effective model complexity and demonstrates its role in explaining double descent and counterintuitive data effects in deep learning.

Findings

Double descent occurs with model size and training epochs.

Increasing data can sometimes harm test performance.

Effective model complexity unifies various double descent phenomena.

Abstract

We show that a variety of modern deep learning tasks exhibit a "double-descent" phenomenon where, as we increase model size, performance first gets worse and then gets better. Moreover, we show that double descent occurs not just as a function of model size, but also as a function of the number of training epochs. We unify the above phenomena by defining a new complexity measure we call the effective model complexity and conjecture a generalized double descent with respect to this measure. Furthermore, our notion of model complexity allows us to identify certain regimes where increasing (even quadrupling) the number of train samples actually hurts test performance.