A Deep Generative Model for Semi-Supervised Classification with Noisy Labels

TL;DR

This paper introduces M-VAE, a semi-supervised deep generative model that explicitly accounts for noisy labels, improving classification performance and providing new theoretical insights into existing models.

Contribution

The paper presents M-VAE, a novel deep generative model that explicitly models label noise, outperforming existing models and offering theoretical insights into semi-supervised learning.

Findings

M-VAE outperforms existing models on noisy label datasets.

Theoretical analysis clarifies the relationship between M-VAE and M1+M2 models.

Explicit noise modeling improves semi-supervised classification accuracy.

Abstract

Class labels are often imperfectly observed, due to mistakes and to genuine ambiguity among classes. We propose a new semi-supervised deep generative model that explicitly models noisy labels, called the Mislabeled VAE (M-VAE). The M-VAE can perform better than existing deep generative models which do not account for label noise. Additionally, the derivation of M-VAE gives new theoretical insights into the popular M1+M2 semi-supervised model.