A Transition Matrix-Based Extended Model for Label-Noise Learning

Haixin Yang, Ruirui Li, Xiangzhong Fang, Yukun Yang, Naihao Wang

Keywords: Transition Matrix, Label-Noise Learning

TL;DR: We propose an extended model for transition matrix that firstly combines it with sparse implicit regularization, enabling the extension of transition matrix methods from class-dependent noise to a broader range of noise scenarios.

Abstract: The transition matrix methods have garnered sustained attention as a class of techniques for label-noise learning due to their simplicity and statistical consistency. However, existing methods primarily focus on class-dependent noise and lack applicability for instance-dependent noise, while some methods specifically designed for instance-dependent noise tend to be relatively complex. To address this issue, we propose an extended model based on transition matrix in this paper, which preserves simplicity while extending its applicability to handle a broader range of noisy data beyond class-dependent noise. The proposed model's consistency and generalization properties are theoretically analyzed under certain assumptions. Experimental evaluations conducted on various synthetic and real-world noisy datasets demonstrate significant improvements over existing transition matrix-based methods. Upon acceptance of our paper, the code will be open sourced.

Primary Area: Deep learning architectures

Submission Number: 18050