Prototype Completion for Few-Shot Learning

TL;DR

This paper introduces a prototype completion framework for few-shot learning that leverages attribute priors and a Gaussian fusion strategy to improve prototype estimation, outperforming existing methods.

Contribution

It proposes a novel meta-learning approach focusing on prototype completion using attribute priors and a fusion strategy, bypassing less effective feature extractor fine-tuning.

Findings

Achieves more accurate class prototypes.

Outperforms existing few-shot learning methods.

Effective in both inductive and transductive settings.

Abstract

Few-shot learning aims to recognize novel classes with few examples. Pre-training based methods effectively tackle the problem by pre-training a feature extractor and then fine-tuning it through the nearest centroid based meta-learning. However, results show that the fine-tuning step makes marginal improvements. In this paper, 1) we figure out the reason, i.e., in the pre-trained feature space, the base classes already form compact clusters while novel classes spread as groups with large variances, which implies that fine-tuning feature extractor is less meaningful; 2) instead of fine-tuning feature extractor, we focus on estimating more representative prototypes. Consequently, we propose a novel prototype completion based meta-learning framework. This framework first introduces primitive knowledge (i.e., class-level part or attribute annotations) and extracts representative features for seen attributes as priors. Second, a part/attribute transfer network is designed to learn to infer the representative features for unseen attributes as supplementary priors. Finally, a prototype completion network is devised to learn to complete prototypes with these priors. Moreover, to avoid the prototype completion error, we further develop a Gaussian based prototype fusion strategy that fuses the mean-based and completed prototypes by exploiting the unlabeled samples. Extensive experiments show that our method: (i) obtains more accurate prototypes; (ii) achieves superior performance on both inductive and transductive FSL settings.