Tackling Noisy Labels With Network Parameter Additive Decomposition

Given data with noisy labels, over-parameterized deep networks suffer overfitting mislabeled data, resulting in poor generalization. The memorization effect of deep networks shows that although the networks have the ability to memorize all noisy data, they would first memorize clean training data, and then gradually memorize mislabeled training data. A simple and effective method that exploits the memorization effect to combat noisy labels is early stopping. However, early stopping cannot distinguish the memorization of clean data and mislabeled data, resulting in the network still inevitably overfitting mislabeled data in the early training stage. In this paper, to decouple the memorization of clean data and mislabeled data, and further reduce the side effect of mislabeled data, we perform additive decomposition on network parameters. Namely, all parameters are additively decomposed into two groups, i.e. , parameters <tex-math notation="LaTeX">w</tex-math> are decomposed as <tex-math notation="LaTeX">w= σ + γ</tex-math> . Afterward, the parameters <tex-math notation="LaTeX">σ</tex-math> are considered to memorize clean data, while the parameters <tex-math notation="LaTeX">γ</tex-math> are considered to memorize mislabeled data. Benefiting from the memorization effect, the updates of the parameters <tex-math notation="LaTeX">σ</tex-math> are encouraged to fully memorize clean data in early training, and then discouraged with the increase of training epochs to reduce interference of mislabeled data. The updates of the parameters <tex-math notation="LaTeX">γ</tex-math> are the opposite. In testing, only the parameters <tex-math notation="LaTeX">σ</tex-math> are employed to enhance generalization. Extensive experiments on both simulated and real-world benchmarks confirm the superior performance of our method.