Concentrating Estimation Attention: Human Prior Constrained Methods for Robust Classification

Zhe Cao, Shuo Yang, Yumeng Liu, Xiaofeng Xu, Hongbin Pei, Yan Huang, Yushu Yu, Ruiheng Zhang

Abstract: In the realm of label noise learning, a potent strategy involves the application of noise transition matrices to foster robust learning processes. Most current research has gained significant success utilizing parameter estimation approaches to generate these matrices when facing instance-dependent noise. Nonetheless, a key drawback of this approach is the diffuse focus of the transition matrix, which can be indiscriminately distracting, thereby ignoring specific locations that are vulnerable to noise flipping. To address this gap, we introduce the Human Attention Constrained Estimation (HACE). This innovative method capitalizes on human cognitive precedents to derive an inter-class affinity matrix. It further refines the estimation of the noise transition matrix by employing our novel Matrix Structure Similarity (MSS) Loss, enabling the matrix estimation module to selectively concentrate on areas frequently affected by noisy flips. This targeted approach addresses the label noise conundrum more effectively and narrows the operational scope significantly. Experiments on three synthetic datasets and a real-world dataset corroborate the robustness and efficiency of our proposed method.