A class-conditional approach to learning with noisy labels

Abstract

[en] The main topic of this work is Learning with Noisy Labels (LNL). It is the field of Machine Learning concerned with training Neural Networks with noisy datasets. In particular, we have studied DivideMix, a method for LNL in the context of Computer Vision. After an extensive research we have discovered that it is unaware of the underlying class-conditional behaviour which consequently produces class imbalances. In this work, we present two class-conditional approaches to DivideMix. With this intent, we study approximate Baye- sian Inference to quantify per-class uncertainty and leverage this extra information to improve the MixMatch step. In addition, we propose a class-aware policy that improves co-divide. Finally, improving DivideMix’s predictive accuracy by up to 0.39% in certain noise settings.