Hallucinating A Cleanly Labeled Augmented Dataset from A Noisy Labeled Dataset Using GAN

Abstract

Noisy labeled learning methods deal with training datasets containing corrupted labels. However, prediction performances of existing methods on small datasets still leave room for improvements. With this objective, in this paper we present a GAN-based method to generate a clean augmented training dataset from a small and noisy labeled dataset. The proposed approach combines noisy labeled learning principles with GAN state-of-the-art techniques. We demonstrate the usefulness of the proposed approach through an empirical study on simple and complex image datasets.

Deep learning with noisy labels: Exploring techniques and remedies in medical image analysis

“…This method was shown to be highly effective in removing label noise and improving the model performance. GANs were used to generate a training dataset with clean labels from an initial dataset with noisy labels by Chiaroni et al (2019).…”

Self-Supervised Learning for Autonomous Vehicles Perception: A Conciliation Between Analytical and Learning Methods

“…• Weakly supervised learning: These techniques can be trained with a partially labeled dataset [6], and eventually with a fraction of corrupted labels [8], [9]. Advantageously, these approaches drastically reduce the need of labeled data.…”

“…Concerning the online self-evaluation, some of the presented systems require a baseline reference analytically obtained [19]. However, if we consider that the analytical processes, considered as ground-truth labeling techniques, are likely to generate some noisy labels, it may be interesting to investigate some future reasearch on how to evaluate this prior noise from the learning model point of view [11], and how to deal with it [9].…”

“…To overcome the former limitation, transfer learning and weakly supervised learning methods have been proposed. Some of them can exploit partially labeled datasets [6], [7], or noisy labeled datasets [8], [9]. Concerning the latter problem, under mild theoretical assumptions on the learning model, we can interpret the predicted outputs.…”

Counter-examples generation from a positive unlabeled image dataset

“…For reasons of expectation linearity, p N would consequently be associated by D to an intermediate label between 0 and 1. A solution may be to extend the proposed framework to this noisy PU learning challenge by drawing on existing asymmetric noisy labeled learning techniques [Chiaroni et al, 2019]. More specifically, by modifying training loss functions, it is possible to enforce G to learn the distribution corresponding to this given intermediate label predicted by D for p N .…”