Idioma:

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

Hu, Wei ; Zhao, QiHao ; Huang, Yangyu ; Zhang, Fan

2020 25th International Conference on Pattern Recognition (ICPR), 2021, p.1882-1889

Sem texto completo

Citações Citado por

Título:
P-DIFF: Learning Classifier with Noisy Labels based on Probability Difference Distributions
Autor: Hu, Wei ; Zhao, QiHao ; Huang, Yangyu ; Zhang, Fan
Assuntos: Benchmark testing ; Computational efficiency ; Neural networks ; Noise measurement ; Pattern recognition ; Task analysis ; Training
É parte de: 2020 25th International Conference on Pattern Recognition (ICPR), 2021, p.1882-1889
Descrição: Learning deep neural network (DNN) classifier with noisy labels is a challenging task because the DNN can easily overfit 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.
Editor: IEEE
Idioma: Inglês

Voltar para lista de resultados