Tackling Instance-Dependent Label Noise via a Universal Probabilistic Model

The drastic increase of data quantity often brings the severe decrease quality, such as incorrect label annotations. It poses a great challenge for robustly training Deep Neural Networks (DNNs). Existing learning methods with noise either employ ad-hoc heuristics or restrict to specific assumptions. However, more general situations, instance-dependent noise, have not been fully explored, scarce studies focus on their corruption process. By categorizing instances into confusing and unconfusing instances, this paper proposes simple yet universal probabilistic model, which explicitly relates noisy labels instances. resultant model can be realized by DNNs, where procedure is accomplished employing novel alternating optimization algorithm. Experiments datasets both synthetic real-world verify proposed method yields significant improvements robustness over state-of-the-art counterparts.