SIGUA: Forgetting May Make Learning with Noisy Labels More Robust

TL;DR

This paper introduces SIGUA, a novel training method that enhances robustness to noisy labels by combining gradient descent on good data with gradient ascent on bad data, effectively reducing undesired memorization.

Contribution

The paper proposes SIGUA, a versatile optimization approach that leverages forgetting undesired memorization to improve learning with noisy labels.

Findings

SIGUA improves robustness of base learning methods against noisy labels.

Experiments show significant performance gains with SIGUA.

The method generalizes across different learning algorithms.

Abstract

Given data with noisy labels, over-parameterized deep networks can gradually memorize the data, and fit everything in the end. Although equipped with corrections for noisy labels, many learning methods in this area still suffer overfitting due to undesired memorization. In this paper, to relieve this issue, we propose stochastic integrated gradient underweighted ascent (SIGUA): in a mini-batch, we adopt gradient descent on good data as usual, and learning-rate-reduced gradient ascent on bad data; the proposal is a versatile approach where data goodness or badness is w.r.t. desired or undesired memorization given a base learning method. Technically, SIGUA pulls optimization back for generalization when their goals conflict with each other; philosophically, SIGUA shows forgetting undesired memorization can reinforce desired memorization. Experiments demonstrate that SIGUA successfully robustifies two typical base learning methods, so that their performance is often significantly improved.