Methods of low data image classification with neural network

Abstract

Image classification has evolved from relying on handcrafted feature engineering to a data-driven deep learning paradigm, with recent breakthroughs extending its capabilities to zeroshot and few-shot learning scenarios. This paper introduces emerging paradigms for low-data scenarios: zero-shot learning methods, including multimodal models (e.g., CLIP) that align visual and textual embeddings for open vocabulary classification and generative frameworks for synthesizing features of unseen classes; and (2) few-shot learning strategies, such as meta-learning (e.g., MAML), metric-based networks (e.g., ProtoNet), and fast adaptation techniques (e.g., Tip-Adapter) that leverage pre-trained knowledge for fast adaptation. While these methods reduce the reliance on labeled data, challenges remain in domain adaptation, fine-grained classification, and computational efficiency.