Heesun Bae, Latent label generation can help improve classification performance degradation when learning with noisy labels, Master's Thesis, Department of Industrial and Systems Engineering, KAIST, 2022 

Abstract

In this paper, we try to solve the performance degradation that happens when the dataset used for training a classifier includes samples with incorrect labels. Several studies have been proposed to solve the label probability distribution distortion due to noisy labels. However, this inference has limits because of lack of information on knowing the authenticity of the label. In result, previous studies of finding the true class distribution, which are either 1) determining the validity of a label or 2) inferring the probability of having a wrong label conditional on the true class by concentrating on the information from only results from classifiers trained with noisy datasets may be hard to converge to the real true distribution. To solve this problem, we propose a method to generate the true class by considering it as latent variables. Specifically, it reflects the available information from the inputs by using a Conditional Variational Autoencoder to generate the true class. However, it may not be suitable to give normal distribution as prior distribution of latent class, since our latent variable will have categorical feature. In this sense, we use the Dirichlet distribution as a prior distribution of our latent variable. Our methodology shows improved performance on several types of mislabeled data.

@masterthesis{Bae:2022,

author = {Heesun Bae},

advisor ={Il-Chul Moon},

title = {Latent label generation can help improve classification performance degradation when learning with noisy labels},

school = {KAIST},

year = {2022}

}