Entropic Regularized Partial Optimal Transport for Efficient Few-Shot Learning

Abstract

Few-shot learning presents the challenge of rapidly and effectively extracting useful information from a limited number of labeled samples. Metric learning methods based on optimal transport theory have become a research hotspot in this area. However, current approaches either suffer from high computational costs, making them impractical for large-scale feature matrix calculations, or they focus solely on full transport between source and target distributions, which can limit performance improvements. In this paper, we propose the Entropic Regularized Partial Optimal Transport (ERPOT) method to address the issue of partial transport. ERPOT supports partial matching between distributions, and the entropic regularization within ERPOT smooths the optimization landscape, reducing the risk of getting trapped in local minima. This characteristic makes the optimization process more reliable and less sensitive to initial conditions. Additionally, we introduce the Receptive Field Cross-Reference (RFC) attention mechanism for feature enhancement. Experimental results demonstrate that our approach achieves significant performance improvements.