Borderline Sample Extraction from a Trained Classifier

Borui Cai, Zihao Johnson Zheng, Longxiang Gao, Yong Xiang

Keywords: Borderline samples, classification, decision boundary

TL;DR: Proposes a Perturbation Minimization method to extract borderline samples from a trained classifier

Abstract: Extracting pseudo samples from a trained classifier helps understand classifier decisions, and extracted samples also can assist downstream tasks like knowledge distillation, continual learning, etc. Existing works mostly focus on extracting exemplary samples, i.e., samples that carry salient features of a class; however, seldom effort has been put into extracting borderline samples that reflect minor differences between two classes. In this paper, we propose a Perturbation Minimization method to extract borderline samples from a trained classifier. Through experiments, we show PM can extract borderline samples, and these samples improve the accuracy in class-incremental learning.

Submission Number: 194