Entropy calibrated prototype embedding for transductive few-shot learning

Abstract

Transductive Few-shot learning aims to generalize to new classes from limited labeled support and all unlabeled query samples. Widely adopted paradigms including prototypical networks and graph-based label propagation. The former classifying queries based on distances to class prototypes, while the latter propagates the labels based on support samples. However, Existing methods typically treat all samples with equal importance, neglect inherent reliabilities, and underutilize prototypes merely as static anchors. This paper proposes Entropy Calibrated Prototype Embedding (ECPE), a novel framework that not only integrates prototypical networks and label propagation methods but also addresses their respective limitations through an iterative refinement strategy. Firstly, we propose the Entropy Calibration (EC), which dynamically assesses sample reliability using prediction entropy to weigh their influence in label propagation. Secondly, Entropy-aware Prototype Embedding (EPE) we proposed treats prototypes as evolving synthetic nodes, iteratively updating them based on calibrated predictions and embedding high-certainty prototypes into the graph.With the iteration of label calibration, entropy-aware prototype embedding, and label propagation, the proposed ECPE enhances classification accuracy and robustness. Extensive experiments demonstrate that ECPE surpasses state-of-the-art performance on three standard Transductive FSL benchmarks. Our source code is published at: https://github.com/gmf-ahu/ECPE.