Robust Classification with Noisy Labels Based on Posterior Learning

Nicola Novello, Andrea M Tonello

Keywords: classification, label noise, f-divergence

Abstract:

Deep learning has shown robustness to label noise under specific assumptions, yet its performance under extremely high noise rates remains a significant challenge.
In this paper, we theoretically demonstrate under what conditions models estimating the posterior probability can achieve high classification accuracy in the presence of extremely strong instance-dependent label noise without performing loss correction approaches.
To estimate the noisy posterior, we propose a class of objective functions derived from the variational representation of the $f$-divergence. Furthermore, we propose two correction methods to achieve robustness when the algorithm is not intrinsically robust to label noise: one method is implemented during the training process, and the other is performed during inference.
Finally, we show the validity of our theoretical results and the effectiveness of the proposed methods on synthetic and real-world label noise settings.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 13599