META-LEARNING FOR FEW-SHOT AND ZERO-SHOT LEARNING: ENABLING RAPID ADAPTATION TO NEW TASKS WITH LIMITED DATA

Abstract

Deep learning systems have achieved remarkable success in computer vision when trained on massive annotated datasets. However, their performance drops drastically when only a few labeled examples are available for new tasks or classes. Humans, by contrast, can generalize from very limited experience. Meta-learning, also known as “learning to learn,” aims to bridge this gap by enabling models to acquire inductive biases that allow rapid adaptation to new tasks with little data. This paper surveys the foundations, key algorithms, and recent advances in meta-learning, few-shot, and zero-shot learning for computer vision. We categorize major approaches metric-based, optimization-based, model-based, and generative and review their contributions to image recognition, detection, and segmentation. We discuss benchmark datasets, challenges, evaluation practices, and open research directions.