Deep Learning Under Instance-Dependent Label Noise: A Tensor Factorization Perspective

Deep learning faces significant challenges with noisy labels, since they substantially deteriorate the generalization performance of the learned classifiers. Among the diverse set of strategies proposed to address this issue, instance-dependent label noise modeling-based approaches are particularly appealing and arguably more realistic. Nonetheless, learning a true-label classifier under instance-dependent modeling is a challenging task. In this work, we propose a novel tensor factorization-based modeling for the instance-dependent label noise, building connections with a block-term tensor decomposition (BTD) model. Our approach utilizes the low-rankness assumption on the latent factors of the BTD model, thereby jointly leveraging both class-specific and instance-specific correlations among the labeling confusions. Additionally, we design an easy-to-train end-to-end learning strategy for the true-label classifier under this framework. Our experiments showcase the effectiveness of the proposed approach.