Confidence Matters: Balanced Pseudo-Label Selection for Semi-Supervised Few-Shot Learning

Semi-Supervised Few-Shot Learning (SSFSL) aims to classify categories given a few support samples and a number of unlabeled samples per class. Most existing methods focus on promoting the pseudo-labeling of unlabeled samples to boost SSFSL performance via more samples with confident pseudo-labels, which is inevitably limited by the scarcity of given support samples for pseudo-labeling training. We realize thatconfidence selectionof pseudo-labeled samples is a promising alternative, for which bothtask-free generalityandtask-specific adaptabilityare crucial to guideaccurateandfine-grainedselection. To this end, we propose a two-step BalAnced pSeudo-label SelectiOn (BASSO) method that explorescertainty entropyto characterize the confidence of pseudo-labels and learns the entropy distributions to guide selection. Specifically, step one rehearses the pseudo-labeling on large task-free samples to model the entropies of high-confidence and low-confidence pseudo-labels as priorgeneral distributions. Step two devises a generator fed with task-specific prototypes to learn theadaptive distributionsfor each task. We further tailorbalancing factorsto balance the two types of distributions in guiding the selection during training episodes, ensuring both accurate and fine-grained selection. We perform confidence selection by casting the certainty entropies of pseudo-labeled samples into the balanced distributions and comparing their probabilities subject to high-confidence and low-confidence pseudo-labels, respectively. The comparison deviations denoting confidence levels are utilized to reweight the selected samples for laterself-training. Extensive experiments on four benchmarks demonstrate that BASSO effectively balances selection accuracy and amount, showcasing competitive performance against state-of-the-art methods. Our code will be available athttps://github.com/Klein-JING/BASSO.