arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.03017cs.LGcs.AI
收藏

少样本学习的当代方法、新兴范式与混合方法综述

A Complete Survey on Contemporary Methods, Emerging Paradigms and Hybrid Approaches for Few-Shot Learning

MetaMind Innovations P.C.(MetaMind创新有限责任公司);Kingston University(金斯顿大学);University of Western Macedonia(西马其顿大学)

Georgios Tsoumplekas, Vladislav Li, Panagiotis Sarigiannidis, Vasileios Argyriou

AI总结:

该综述全面梳理少样本学习(FSL)的成熟方法与最新进展,扩展分类法涵盖上下文学习、神经过程等新兴范式及混合方法,并探讨应用、挑战与未来方向。

摘要:

尽管深度学习取得了广泛成功,但其对海量数据和大量训练的强烈需求,使其在数据稀缺的各种实际应用中变得不切实际。近年来,少样本学习(Few-Shot Learning, FSL)作为一种学习范式应运而生,旨在通过利用先验知识实现对新学习任务的快速适应,从而解决这些局限性。由于FSL的特性与深度学习对数据密集型需求高度互补,其在过去几年中得到了显著发展。本综述全面概述了FSL领域中既有的成熟方法以及最新进展。所提出的分类法在先前分类的基础上进行了扩展,纳入了新兴的FSL范式,如上下文学习(in-context learning),以及元学习范式中针对FSL的新类别,包括神经过程(neural processes)和概率元学习(probabilistic meta-learning)。此外,通过讨论将FSL扩展到通常考察的监督学习设置之外的混合FSL方法,提供了对FSL的整体概述。本综述还探讨了FSL在各个领域的多样化应用。最后,讨论了塑造该领域的最新趋势、突出挑战以及有前景的未来研究方向。

英文摘要:

Despite the widespread success of deep learning, its intense requirements for vast amounts of data and extensive training make it impractical for various real-world applications where data is scarce. In recent years, Few-Shot Learning (FSL) has emerged as a learning paradigm that aims to address these limitations by leveraging prior knowledge to enable rapid adaptation to novel learning tasks. Due to its properties that highly complement deep learning's data-intensive needs, FSL has seen significant growth in the past few years. This survey provides a comprehensive overview of both well-established methods as well as recent advancements in the FSL field. The presented taxonomy extends previously proposed ones by incorporating emerging FSL paradigms, such as in-context learning, along with novel categories within the meta-learning paradigm for FSL, including neural processes and probabilistic meta-learning. Furthermore, a holistic overview of FSL is provided by discussing hybrid FSL approaches that extend FSL beyond the typically examined supervised learning setting. The survey also explores FSL's diverse applications across various domains. Finally, recent trends shaping the field, outstanding challenges, and promising future research directions are discussed.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑