零样本学习进展研究

Review of Zero-Shot Learning

机器学习是人工智能领域发展最迅速的一个分支之一,传统的机器学习方法和深度学习大都需要大量人工标注的训练数据才能发挥作用.然而,现实世界的物体种类繁多且其数量在不断增长,人工标注训练数据就变成了一项极其繁琐冗杂的工作,零样本学习的提出极大地缓解了这种情况.在零样本学习中,训练集和测试集的类别的交集是空集,因此需要在二者之间通过实现知识的迁移来完成学习,从而使得在训练集上训练得到的模型能够识别测试集上输入示例的类别标签.不同于其他大部分机器学习技术需要保证训练集包含测试集,零样本学习的原理从本质意义上让计算机模仿了人类在学习时的推理模式,使得计算机能够识别新事物.本文梳理了零样本学习的研究进展,首先概述了零样本学习的定义及其相关领域,然后重点归纳了零样本学习的发展过程,包括其基本模型及改进,存在的关键难点以及解决方式,最后探讨了零样本学习的研究现状及其未来的发展方向.

Machine learning is one of the fastest-growing branches in the field of artificial intelligence.Most of the traditional machine learning methods and deep learning require a large amount of manually annotated training data to function.However,on the on hand,there are many types of objects in the real world and their number is increasing,so manually annotated training data must be updated.This is an extremely tedious work.Therefore,the proposal of zero-shot learning has greatly alleviated this situation.In zero-shot learning,the intersection of the categories of the training set and the test set is an empty set,so it is necessary to complete the learning process by implementing knowledge transfer between the two sets,so that the model trained on the training set can recognize the category labels of the input examples in the test set.Unlike most other machine learning technologies that need to ensure that the training set contains the test set,the principle of zero-shot learning essentially allows the computer to imitate the reasoning mode of humans during learning,so that the computer can recognize new things.This article combs the research progress of zero-shot learning.First,it outlines the definition of zero-shot learning and its related fields,and then focuses on the development process of zero-shot learning,including its basic models and their improvements,key difficulties and solutions.Finally,based on the research status of zero-shot learning,its future direction is discussed.