TrustNet: Learning from Trusted Data Against (A)symmetric Label Noise

Amirmasoud Ghiassi,Robert Birke,Lydia Y.Chen

doi:10.1145/3492324.3494166

Abstract

Robustness to label noise is a critical property for weakly-supervised classifiers trained on massive datasets. Robustness to label noise is a critical property for weakly-supervised classifiers trained on massive datasets. In this paper, we first derive analytical bound for any given noise patterns. Based on the insights, we design TrustNet that first adversely learns the pattern of noise corruption, being it both symmetric or asymmetric, from a small set of trusted data. Then, TrustNet is trained via a robust loss function, which weights the given labels against the inferred labels from the learned noise pattern. The weight is adjusted based on model uncertainty across training epochs. We evaluate TrustNet on synthetic label noise for CIFAR-10 and CIFAR-100, and real-world data with label noise, i.e., Clothing1M. We compare against state-of-the-art methods demonstrating the strong robustness of TrustNet under a diverse set of noise patterns.

Highlights

IntroductionThe majority of deep networks robust against dirty labels focuses on synthetic label noise, which can be symmetric or asymmetric
We aim to show the effectiveness of TrustNet via testing accuracy on diverse and challenging noise patterns
Motivated by the disparity of label noise patterns studied in the prior state-of-the-art methods, we first derive the analytical understanding of synthetic and real-world noise, i.e., how testing accuracy degrades with noise ratios and patterns

Summary

Introduction

The majority of deep networks robust against dirty labels focuses on synthetic label noise, which can be symmetric or asymmetric. The former case [2] assumes noise labels can be corrupted into any other classes with equal probability. Patterns of noisy labels observed from real-life big data sets, e.g., Clothing1M [38], exhibit high percentages of label noise and more complicated patterns mixing symmetric and asymmetric noises. There is disagreement among related work on which noise patterns are more detrimental and difficult to defend against for regular networks [21, 34]

Objectives

Methods

Discussion

Conclusion