克服小样本学习中灾难性遗忘方法研究

OVERCOMING CATASTROPHIC FORGETTING IN FEW-SHOT LEARNING

针对小样本学习在识别新类别时会出现灾难性遗忘的问题,提出一种小样本学习中克服灾难性遗忘的方法。结合卷积神经网络识别模型提取图片特征,引用注意力机制设计分类权重生成器,使新类权重的生成基于基类权重。通过基于皮尔森相似度的识别模型计算新类特征与基类图片分类权重之间的相似度,判断新类图像的类别。在三种数据集进行实验,结果表明:该方法使小样本图像分类的精度得到了一定程度的提升,同时不会牺牲基类的识别准确度,克服了灾难性遗忘。

Few-shot learning can cause catastrophic forgetting when identifying new categories.This paper studies the method of overcoming catastrophic forgetting in few-shot learning.Combining the recognition model of convolutional neural network,the image features were extracted,and the attentional mechanism was used to design the classification weight generator,so that the new class weights were generated based on the base class weights.It calculated the similarity between the new class features and the classification weight of the base class image through the recognition model based on Pearson similarity,and then judged the category of the new class image.Three data sets were used for experiments.The results show that the proposed method improves the accuracy of few-shot image classification to a certain extent.And it does not sacrifice the recognition accuracy of the base class,which overcomes the catastrophic forgetting to some extent.