Joint Optimization Framework for Learning with Noisy Labels

TL;DR

This paper introduces a joint optimization framework that simultaneously learns neural network parameters and estimates true labels, effectively correcting noisy labels during training to improve image classification performance.

Contribution

It presents a novel method that alternates between updating network parameters and label estimates, outperforming existing techniques on noisy datasets.

Findings

Significant performance improvement on noisy CIFAR-10 and Clothing1M datasets.

Outperforms state-of-the-art methods in handling noisy labels.

Effective label correction during training enhances model robustness.

Abstract

Deep neural networks (DNNs) trained on large-scale datasets have exhibited significant performance in image classification. Many large-scale datasets are collected from websites, however they tend to contain inaccurate labels that are termed as noisy labels. Training on such noisy labeled datasets causes performance degradation because DNNs easily overfit to noisy labels. To overcome this problem, we propose a joint optimization framework of learning DNN parameters and estimating true labels. Our framework can correct labels during training by alternating update of network parameters and labels. We conduct experiments on the noisy CIFAR-10 datasets and the Clothing1M dataset. The results indicate that our approach significantly outperforms other state-of-the-art methods.