Uncertainty-aware self-training with expectation maximization basis transformation

TL;DR

This paper introduces an uncertainty-aware self-training framework using expectation maximization and basis transformation to improve pseudo-label quality, leading to better performance in image classification and segmentation.

Contribution

It proposes a novel self-training method that incorporates uncertainty estimation through EM and basis extraction, addressing over-confidence issues in pseudo-labeling.

Findings

Achieves 1-3% accuracy improvement over existing confidence-aware self-training methods.

Effectively estimates and filters uncertainty to refine pseudo-labels.

Demonstrates superior performance on image classification and semantic segmentation tasks.

Abstract

Self-training is a powerful approach to deep learning. The key process is to find a pseudo-label for modeling. However, previous self-training algorithms suffer from the over-confidence issue brought by the hard labels, even some confidence-related regularizers cannot comprehensively catch the uncertainty. Therefore, we propose a new self-training framework to combine uncertainty information of both model and dataset. Specifically, we propose to use Expectation-Maximization (EM) to smooth the labels and comprehensively estimate the uncertainty information. We further design a basis extraction network to estimate the initial basis from the dataset. The obtained basis with uncertainty can be filtered based on uncertainty information. It can then be transformed into the real hard label to iteratively update the model and basis in the retraining process. Experiments on image classification and semantic segmentation show the advantages of our methods among confidence-aware self-training algorithms with 1-3 percentage improvement on different datasets.