CCT: A Cyclic Co-Teaching Approach to Train Deep Neural Networks With Noisy Labels
15 mai. 2025·,,,·
1 minutos de leitura
Jorge K. S. Kamassury
Henrique Pickler
Filipe R. Cordeiro
Danilo Silva
Resumo
Deep neural networks are highly susceptible to memorizing incorrect labels, compromising generalization on real-world datasets with inaccurate annotations. Among existing approaches, Co-Teaching trains two models in parallel to identify potentially noisy samples through cross-selection, but still suffers from error accumulation and overfitting. We propose Cyclic Co-Teaching (CCT), which mitigates these limitations through periodic modulations of the learning rate and sample retention, establishing an alternating dynamic between specialization and consolidation phases. We also introduce a two-step univariate optimization for hyperparameter tuning. CCT consistently outperforms state-of-the-art methods on synthetic (CIFAR-10, CIFAR-100, Tiny-ImageNet) and real-world (Animal-10N, Food-101N, Clothing1M) benchmarks, particularly under high-noise scenarios.
Tipo
Publicação
IEEE Access, 13, 43843–43860
Propomos Cyclic Co-Teaching (CCT), um método de treinamento robusto para redes neurais profundas em bases com rótulos ruidosos. O método combina modulações cíclicas da taxa de aprendizado e retenção de amostras, criando uma dinâmica alternada entre fases de especialização (aprendizado intensivo) e consolidação (estabilização). Experimentos em bases sintéticas (CIFAR-10/100, Tiny-ImageNet) e reais (Animal-10N, Food-101N, Clothing1M) mostram ganhos consistentes sobre o estado da arte, especialmente em cenários de alta taxa de ruído.