A Meta Approach to Defend Noisy Labels by the Manifold Regularizer PSDR

TL;DR

This paper introduces PSDR, a manifold regularizer that leverages data augmentation to improve deep neural network robustness against noisy labels by utilizing all samples effectively.

Contribution

The paper proposes PSDR, a novel manifold regularizer that penalizes divergence between similar inputs' outputs, enhancing robustness to noisy labels without discarding data.

Findings

Significantly improves state-of-the-art results on benchmark datasets.

Effectively utilizes all samples, including noisy ones.

Easily implementable on various DNN architectures.

Abstract

Noisy labels are ubiquitous in real-world datasets, which poses a challenge for robustly training deep neural networks (DNNs) since DNNs can easily overfit to the noisy labels. Most recent efforts have been devoted to defending noisy labels by discarding noisy samples from the training set or assigning weights to training samples, where the weight associated with a noisy sample is expected to be small. Thereby, these previous efforts result in a waste of samples, especially those assigned with small weights. The input $x$ is always useful regardless of whether its observed label $y$ is clean. To make full use of all samples, we introduce a manifold regularizer, named as Paired Softmax Divergence Regularization (PSDR), to penalize the Kullback-Leibler (KL) divergence between softmax outputs of similar inputs. In particular, similar inputs can be effectively generated by data augmentation. PSDR can be easily implemented on any type of DNNs to improve the robustness against noisy labels. As empirically demonstrated on benchmark datasets, our PSDR impressively improve state-of-the-art results by a significant margin.