Semantic Pruning for Single Class Interpretability

Kamila Abdiyeva, Martin Lukac, Kanat Alimanov

Keywords: deep learning, semantic pruning, filter correlation

TL;DR: Semantic Pruning for Filter Interpretability

Abstract: Convolutional Neural Networks (CNN) have achieved state-of-the-art performance in different computer vision tasks, but at a price of being computationally and power intensive. At the same time, only a few attempts were made toward a deeper understanding of CNNs. In this work, we propose to use semantic pruning technique toward not only CNN optimization but also as a way toward getting some insight information on convolutional filters correlation and interference. We start with a pre-trained network and prune it until it behaves as a single class classifier for a selected class. Unlike the more traditional approaches which apply retraining to the pruned CNN, the proposed semantic pruning does not use retraining. Conducted experiments showed that a) for each class there is a pruning ration which allows removing filters with either an increase or no loss of classification accuracy, b) pruning can improve the interference between filters used for classification of different classes c) effect between classification accuracy and correlation between pruned filters groups specific for different classes.