A Study on Metric-Based and Initialization-Based Methods for Few-Shot Image Classification

Abstract

Few-shot learning (FSL), or learning to generalize using few training data samples, is a particularly challenging problem in machine learning. This paper discusses various state-of-the-art distance metric-based and initialization-based FSL methods. It also gives a background on the meta-learning framework employed by many of the discussed models to generalize to novel classification tasks after training on multiple training tasks. We also discuss other techniques, such as whole-class classification, that have produced better results than meta-learning for metric-based methods. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.