Associative Alignment for Few-shot Image Classification

TL;DR

This paper introduces associative alignment strategies to improve few-shot image classification by leveraging related base data, significantly enhancing accuracy across multiple datasets and tasks.

Contribution

It proposes novel metric-learning and adversarial alignment methods that expand effective training data, enabling better fine-tuning for few-shot learning scenarios.

Findings

Achieved up to 6.2% accuracy improvement in 5-shot learning.

Demonstrated effectiveness across four datasets and three backbone architectures.

Outperformed state-of-the-art methods in various few-shot classification tasks.

Abstract

Few-shot image classification aims at training a model from only a few examples for each of the "novel" classes. This paper proposes the idea of associative alignment for leveraging part of the base data by aligning the novel training instances to the closely related ones in the base training set. This expands the size of the effective novel training set by adding extra "related base" instances to the few novel ones, thereby allowing a constructive fine-tuning. We propose two associative alignment strategies: 1) a metric-learning loss for minimizing the distance between related base samples and the centroid of novel instances in the feature space, and 2) a conditional adversarial alignment loss based on the Wasserstein distance. Experiments on four standard datasets and three backbones demonstrate that combining our centroid-based alignment loss results in absolute accuracy improvements of 4.4%, 1.2%, and 6.2% in 5-shot learning over the state of the art for object recognition, fine-grained classification, and cross-domain adaptation, respectively.