Few-Shot Learning
小样本学习(Few-Shot Learning)是一种机器学习范式,旨在使模型仅凭极少量标注样本(通常为1至数十个)即可完成对新任务或新类别的识别与泛化。其核心思想源于元学习(Meta-Learning),通过在多个辅助任务上训练,使模型习得跨任务迁移的通用能力。常见方法包括基于度量的方法、基于优化的方法和基于模型的方法,广泛应用于图像识别、自然语言处理等数据稀缺场景。
Core Facts
Few-Shot Learning refers to providing a small number of examples (typically 2–8) in the prompt, allowing the model to complete new tasks through analogical reasoning.
Few-shot prompting activates in-context learning capabilities by providing examples.
All Facts (6)
Few-Shot Learning refers to providing a small number of examples (typically 2–8) in the prompt, allowing the model to complete new tasks through analogical reasoning.
80%VerifiedFew-shot prompting activates in-context learning capabilities by providing examples.
65%Unverified提示词工程已发展出思维链(Chain-of-Thought)、少样本学习(Few-Shot Learning)、角色扮演(Role Prompting)等多种成熟范式
90%Unverified小样本学习(Few-Shot Learning)的主流方法包括基于度量学习的Siamese Network、基于元学习的MAML,以及结合预训练大模型的提示微调
85%Unverified使用少样本提示词模板时,2-5个高质量、有代表性的示例胜过大量低质量示例,示例的多样性和边界覆盖度比数量更重要
75%Unverified少样本方式在格式化输出、分类任务和规则遵循方面的幻觉率可降低40%-70%
60%