Novel Norms for Pruning Convolutional Neural Networks

Abstract

Convolutional Neural Networks (CNNs) for object detection in images can have millions of parameters, leading to a large memory footprint and relatively long inference times. These parameters are grouped into objects called filters, and structured pruning is the process of removing a proportion of filters from a CNN without significantly reducing its accuracy. The manner in which filters are chosen affects the performance of the model. One method for selecting which filters to prune is to use a norm function to assign a numeric value to each filter and prune the ones with the smallest value, since those filters are typically the least important. Existing works only use a handful of norms for pruning. In this work, we propose a method to evaluate the effectiveness of a wide range of norms and we develop a novel norm family for pruning. We compare the results of our experiments and show that our norms are more effective for pruning than conventional norms for the YOLOv10n model. Our approach for evaluating the effectiveness of norms for pruning can be used by researchers with other CNN models, such as other Ultralytics YOLO models or CNN models for applications other than object detection, such as image classification or semantic parsing. The researchers may experiment with the norms presented in this thesis as well as with norms they may define themselves.