类特征增强元学习算法  

Meta-learning with Class Feature Augmentation

作　　者：蔡奇 李凡长[1] CAI Qi;LI Fan-zhang(School of Computer Science and Technology,Soochow University,Suzhou 215006,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《小型微型计算机系统》2022年第2期225-230,共6页Journal of Chinese Computer Systems

基　　金：国家重点研发计划“变革性技术关键科学问题”重点专项项目(2018YFA0701700,2018YFA0701701)资助.

摘　　要：深度学习取得了巨大成就,尤其是在计算机视觉领域,已经接近人类水平.但是这些成果大多依赖于巨大规模的训练数据,在面对数据量较小的情况时,往往产生严重的过度拟合和灾难性的遗忘.最近的研究表明,采用元学习的方法可以解决此问题.元学习模型由特征提取与分类算法两个模块组成,目前大部分研究都关注于设计合适的分类算法,忽视了特征的重要性.本文认为每个类都有自身独有的一些类特征,这些特征更明显的图像更易于识别.因此提出了类特征增强的方法,借助相似类来获得类特征,在特征的层面上对图像进行增强,使提取后的特征更有利于当前的分类任务.类特征增强是一个通用框架,可以与大多数元学习方法相结合.本文将其与常用的原型网络相结合,在Mini-ImageNet数据集上取得了更好的实验效果.Deep learning achieves great achievements,especially in the field of computer vision,which is already close to human level.However,most of these achievements rely on huge-scale training data.When trained with small data,it suffers from over-fitting and catastrophic forgetting.Recent studies show that meta-learning can solve such problems.Most meta-learning models consist of two modules:feature extraction and classification algorithm.At present,researches focus on the design of appropriate classification algorithm,ignoring the importance of features.We think each class has its own unique class features,image with these features is more recognizable.We propose Class Feature Augmentation,using similar classes to get class features,augmenting the images at feature level.Therefore,the extracted features are more conducive to the current classification task.Class Feature Augmentation is a general framework that can be combined with most meta-learning methods.In this paper,we combine it with Prototype Network and achieve better result on Mini-ImageNet.

关 键 词：元学习 特征增强 小样本学习 深度学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]