A survey of intelligent algorithms in few-shot learning

The purpose of few-shot recognition is to tackle the challenge of identifying new categories when only limited labeled samples are available. This survey summarizes four major branches of few-shot learning—meta-learning, transfer learning, data augmentation and generative models, and graph neural networks—while tracing their development, strengths, and limitations. Building on current trends and emerging opportunities in few-shot recognition, this work further discusses potential research directions. Future progress in this field is expected to hinge on three key avenues: cross-domain few-shot recognition, methods that integrate multimodal learning with self-supervised pretraining, and lightweight few-shot learning frameworks. These directions provide valuable reference points and guidance for the advancement of intelligent algorithms under data-scarce conditions.