Zero-Shot Learning

Zero-Shot Learning (ZSL) refers to the model's ability to recognize certain categories that it has not encountered during the training process. Its core objective is to achieve effective classification and recognition on categories that were unknown during the supervised learning phase. In modern NLP, language models can evaluate downstream tasks without fine-tuning, significantly enhancing the model's generalization ability and application value. ZSL achieves inference on unseen categories by learning a mapping from the image feature space to the semantic space, or through nonlinear multimodal embeddings. Benchmark datasets such as aPY, AwA, and CUB have provided crucial support for ZSL research.