P-DIFF: Learning Classifier with Noisy Labels based on Probability Difference Distributions

TL;DR

This paper introduces P-DIFF, a simple yet effective training method for deep neural networks that mitigates the impact of noisy labels by using probability difference distributions to re-weight samples, improving robustness without prior noise knowledge.

Contribution

The paper proposes P-DIFF, a novel training paradigm that implicitly estimates sample cleanliness and enhances DNN robustness against label noise without needing noise rate prior.

Findings

P-DIFF outperforms state-of-the-art methods on benchmark datasets.

It effectively reduces overfitting to noisy labels.

The method works without prior noise rate knowledge.

Abstract

Learning deep neural network (DNN) classifier with noisy labels is a challenging task because the DNN can easily over-fit on these noisy labels due to its high capability. In this paper, we present a very simple but effective training paradigm called P-DIFF, which can train DNN classifiers but obviously alleviate the adverse impact of noisy labels. Our proposed probability difference distribution implicitly reflects the probability of a training sample to be clean, then this probability is employed to re-weight the corresponding sample during the training process. P-DIFF can also achieve good performance even without prior knowledge on the noise rate of training samples. Experiments on benchmark datasets also demonstrate that P-DIFF is superior to the state-of-the-art sample selection methods.