<ul><li>This paper discusses long-tailed semi-supervised learning (LTSSL) with distribution mismatch, where the class distribution of labeled and unlabeled training data differs.</li><li>Existing methods use auxiliary classifiers (experts) to generate pseudo-labels, but the expertise of these experts is not fully utilized.</li><li>A dynamic expert assignment module is proposed to estimate sample class membership and assign suitable experts based on intervals, improving pseudo-label quality during training and predictions during testing.</li><li>The study shows integrating different experts' strengths lowers generalization error and introduces a multi-depth feature fusion module to address model bias effectively, validated through experiments on CIFAR-10-LT, STL-10-LT, and SVHN-LT datasets.</li></ul>

A Square Peg in a Square Hole: Meta-Expert for Long-Tailed Semi-Supervised Learning

Discover more