Probabilistic Decoupling of Labels in Classification

TL;DR

This paper introduces a probabilistic framework for various complex classification tasks, enabling the inference of true class distributions from noisy or incomplete labels through variational optimization.

Contribution

It presents a unified probabilistic approach that models label-class transitions to improve classification in semi-supervised, positive-unlabeled, and noisy-label scenarios.

Findings

Effective inference of class distributions from noisy labels

Unified approach applicable to multiple non-standard classification tasks

Improved accuracy in label-distribution estimation

Abstract

In this paper we develop a principled, probabilistic, unified approach to non-standard classification tasks, such as semi-supervised, positive-unlabelled, multi-positive-unlabelled and noisy-label learning. We train a classifier on the given labels to predict the label-distribution. We then infer the underlying class-distributions by variationally optimizing a model of label-class transitions.