Provable Benefit of Cutout and CutMix for Feature Learning

TL;DR

This paper provides a theoretical analysis showing that Cutout and CutMix augmentations enable neural networks to learn a broader range of features, with CutMix offering the most comprehensive feature learning and highest test accuracy.

Contribution

It offers the first theoretical understanding of how Cutout and CutMix improve feature learning in neural networks, especially regarding rare and noisy features.

Findings

Cutout learns low-frequency features that vanilla training cannot.

CutMix learns rarer features that Cutout cannot.

CutMix achieves the highest test accuracy among the methods.

Abstract

Patch-level data augmentation techniques such as Cutout and CutMix have demonstrated significant efficacy in enhancing the performance of vision tasks. However, a comprehensive theoretical understanding of these methods remains elusive. In this paper, we study two-layer neural networks trained using three distinct methods: vanilla training without augmentation, Cutout training, and CutMix training. Our analysis focuses on a feature-noise data model, which consists of several label-dependent features of varying rarity and label-independent noises of differing strengths. Our theorems demonstrate that Cutout training can learn low-frequency features that vanilla training cannot, while CutMix training can learn even rarer features that Cutout cannot capture. From this, we establish that CutMix yields the highest test accuracy among the three. Our novel analysis reveals that CutMix training makes the network learn all features and noise vectors "evenly" regardless of the rarity and strength, which provides an interesting insight into understanding patch-level augmentation.