AggMatch: Aggregating Pseudo Labels for Semi-Supervised Learning

TL;DR

AggMatch is a semi-supervised learning framework that improves pseudo label quality by aggregating information from similar instances and using a confidence-aware queue, leading to better performance on benchmarks.

Contribution

The paper introduces AggMatch, a novel SSL method that refines pseudo labels through aggregation and confidence measures, addressing noise issues in pseudo labeling.

Findings

AggMatch outperforms recent SSL methods on standard benchmarks.

Aggregation improves pseudo label accuracy and stability.

The confidence measure enhances pseudo label reliability.

Abstract

Semi-supervised learning (SSL) has recently proven to be an effective paradigm for leveraging a huge amount of unlabeled data while mitigating the reliance on large labeled data. Conventional methods focused on extracting a pseudo label from individual unlabeled data sample and thus they mostly struggled to handle inaccurate or noisy pseudo labels, which degenerate performance. In this paper, we address this limitation with a novel SSL framework for aggregating pseudo labels, called AggMatch, which refines initial pseudo labels by using different confident instances. Specifically, we introduce an aggregation module for consistency regularization framework that aggregates the initial pseudo labels based on the similarity between the instances. To enlarge the aggregation candidates beyond the mini-batch, we present a class-balanced confidence-aware queue built with the momentum model, encouraging to provide more stable and consistent aggregation. We also propose a novel uncertainty-based confidence measure for the pseudo label by considering the consensus among multiple hypotheses with different subsets of the queue. We conduct experiments to demonstrate the effectiveness of AggMatch over the latest methods on standard benchmarks and provide extensive analyses.