Relation fusion propagation network for transductive few-shot learning

Abstract

Previous graph-based meta-learning approaches have explored pairwise feature similarity to learn instance-level relations of samples, however, the gap between the sample relations in feature and label spaces is often ignored. It is empirically observed that instances with different labels may display considerable similarity in visual characteristics, making it challenging to distinguish between them in the feature space. To this end, we propose a dual-branch Relation Fusion Propagation Network (RFPN) for transductive few-shot learning, which explicitly models both feature and label relations across support-query pairs. Specifically, we design a Relation Fusion Block (RFB) to fuse instance-level and class-level relations, thus obtaining more robust fusion-level relations to guide feature propagation. In addition to the feature propagation branch, we encode the label relations and construct a label shortcut branch for label propagation. To alleviate the error accumulation during propagation, which is caused by uncertain pseudo-labels for query samples, we propose a Label Shortcut Mechanism (LSM) to progressively update the sample relations with the initial labels. Our full method is plug-and-play and can be easily applied in existing graph-based approaches for transductive few-shot learning. Extensive experiments demonstrate that our proposed RFPN yields significant improvements over the baselines and achieves promising performance on four popular few-shot classification benchmarks.