关键词: 小样本学习, 对比学习, 孪生神经网络, 滑动池化

Abstract: Aiming at the problems existing in few-shot learning at present, we designed a new network structure and its training method to improve the few-shot learning. The  convolution network and multi-scale slide pooling method were used to enhance feature extraction in the feature embedding part of the network. The main structure  of the networks was the Siamese network  to facilitate learning semantics from small sample data through comparison between samples. The training method  of the framework adopted nested level parameter updating to ensure the stability of convergence. Compared with the common visual model and 
few-shot learning methods, the experimental results  on two classical few-shot learning datasets show that the method significantly improves the  accuracy of  few-shot learning, and  can be used as a solution  under the condition of insufficient sample.

Key words: few-shot learning, contrastive learning, Siamese network, slide pooling