A robust classifier for Noise-corruption Learning
Linchang Zhao, Hao Wei, Mu Zhang, Ruiping Li, Qianbo Li, Hongming Cai · 2023
Deep neural networks (DNNs) have achieved great success in various applications across many disciplines. However, their superior performance is susceptible to training set bias and noise label corruption. In order to overcome the overfitting of deep network models to corrupted labels, in addition to various regularization techniques and example re-weighting algorithms, teacher-student models based on the knowledge distillation paradigm are also a popular approach to solve these problems. However, these methods require careful tuning of additional hyperparameters, such as example mining plans and regularization hyperparameters. Unlike previous re-weighting methods, this paper proposes a novel meta-learning algorithm based on deep transfer hybrid model for training example weight allocation. The method combines the RN50STN model designed based on the spatial transformer network (STN) and pre-trained ResNet50, resulting in good performance in a wide range of noisy label corruption.