FLAP: A Semi-Supervised Few-Shot Learning Algorithm via Pseudo-Labeling

Abstract Few-shot learning addresses the challenge of training models with limited samples, striving to achieve robust classification performance with only a few examples, particularly for novel objects. One challenging aspect of few-shot learning is the cross-domain learning scenario, characterized by significant differences in data distribution between the target and source domains. To bridge this gap and improve domain adaptation, we introduce unlabeled samples from new categories within the target domain during training. Our proposed semi-supervised few-shot classification algorithm, FLAP, employs clustering methods to predict labels for unlabeled samples, enabling it to adapt to new unlabeled categories. We have also integrated a distance-based loss to consider clustering characteristics during the training of feature encoders and classifiers. This approach, combined with the symmetric cross-entropy loss, ensures robustness against incorrect labels introduced during pseudo-labeling. Experimental studies validate the effectiveness of our method, demonstrating significant accuracy improvements across various cross-domain and domain-constrained few-shot learning tasks.