Few shot learning for multi-class classification based on nested ensemble DSVM

Abstract Few shot learning (FSL) is a challenge task because a large amount of labeled samples are difficult to acquire due to privacy, ethic issues, safety concerns. Nevertheless, present approaches for multi-class classification usually require massive labeled samples. To bridge this gap, we propose a novel approach named the Nested Ensemble Deep Support Vector Machine (NE-DSVM). In this method, we firstly assign three different single-kernel functions to the deep support vector (DSVM) and ensemble a base classifier in the inner layer. Then, in the outer layer, we use the “one-to-rest” (OvR) strategy to convert the multi-class classification into a binary-class classification. In the last, we combine the AdaBoost framework to complete the classification task. To evaluate the effectiveness of the proposed method, we build a through wall human being detection system and conduct five-fold cross validation experiments on three datasets. We further compare it with SVM, DSVM, AdaBoost-SVM algorithm on the same training and testing data. The experimental results show that the proposed NE-DSVM algorithm lead to encouraging results for few shot learning.