Error Correction for Semi-Supervised Classification Based on Fix Match

The College of Engineering, Computing and Cybernetics, The Australian National University (ANU), Canberra, ACT 2601, Australia

^* Corresponding author: u7545724@anu.edu.au

Abstract

Semi-supervised learning (SSL) leverages unlabeled data to support model training, thereby improving model accuracy. However, most existing SSL methods rely heavily on unlabeled data for model correction while often overlooking the potential error correction capabilities of labeled data. In this paper, we propose Correction FixMatch, which harnesses the error correction potential of labeled data to achieve higher accuracy on the test set. Correction FixMatch is based on the FixMatch model but introduces a novel error-correction mechanism utilizing labeled data. In particular, a training set and a correction set are separated from the labeled dataset. The model is initially trained using the training set, and after a predefined number of steps, the correction set is employed to refine the model through error correction. According to experimental results, the suggested approach performs noticeably better in terms of accuracy than the original FixMatch model under identical testing conditions. This method opens up new possibilities for improving semi-supervised learning models.