Generalizing and Classifying From Few Samples: A Comprehension of Approaches to Few-Shot Visual Learning

Abstract

Unlike traditional machine learning techniques, few-shot learning (FSL) represents a paradigm aimed at acquiring new tasks from just a handful of labeled examples. The challenge in FSL lies in its requirement for models to generalize effectively from a small dataset to previously unseen examples. Various approaches have been developed for FSL, encompassing techniques such as metric learning, meta-learning, and hybrid methods, among others. These approaches have found success in numerous computer vision tasks, including image and video classification, object detection, object segmentation, robotics, natural language processing, and various real-world applications such as medical diagnosis and self-driving cars. This comprehensive survey offers an in-depth exploration of recent advancements and the current state-of-the-art in FSL. The study presents a thorough examination of different FSL approaches, categorizing them primarily into meta-learning and non-meta-learning methods. It also delves into benchmark datasets for FSL, highlights existing research challenges, and explores the diverse applications of FSL. Furthermore, the survey identifies and discusses open research challenges within the field of FSL.