Rethinking the Value of Labels for Instance-Dependent Label Noise Learning

Label noise widely exists in large-scale datasets and significantly degenerates the performances of deep learning algorithms. Due to the non-identifiability of the instance-dependent noise transition matrix, most existing algorithms address the problem by assuming the noisy label generation process to be independent of the instance features. Unfortunately, noisy labels in real-world applications often depend on both the true label and the features. In this work, we tackle instance-dependent label noise with a novel deep generative model that avoids explicitly modeling the noise transition matrix. Our algorithm leverages casual representation learning and simultaneously identifies the high-level content and style latent factors from the data. By exploiting the supervision information of noisy labels with structural causal models, our empirical evaluations on a wide range of synthetic and real-world instance-dependent label noise datasets demonstrate that the proposed algorithm significantly outperforms the state-of-the-art counterparts.

Towards Learning Causal Representations from Multi-Instance Bags

Although humans can easily identify the object of interest from groups of examples using group-level labels, most of the existing machine learning algorithms can only learn from individually labeled examples. Multi-instance learning (MIL) is a type of weakly supervised learning that deals with objects represented as groups of instances, and is theoretically capable of predicting instance labels from group-level supervision. Unfortunately, most existing MIL algorithms focus on improving the performances of group label predictions and cannot be used to accurately predict instance labels. In this work, we propose the TargetedMIL algorithm, which learns semantically meaningful representations that can be interpreted as causal to the object of interest. Utilizing the inferred representations, TargetedMIL excels at instance label predictions from group-level labels. Qualitative and quantitative evaluations on various datasets demonstrate the effectiveness of TargetedMIL.