FALF ConvNets: Fatuous auxiliary loss based filter-pruning for efficient deep CNNs

作者:

Highlights:

摘要

Obtaining efficient Convolutional Neural Networks (CNNs) are imperative to enable their application for a wide variety of tasks (classification, detection, etc.). While several methods have been proposed to solve this problem, we propose a novel strategy for solving the same that is orthogonal to the strategies proposed so far. We hypothesize that if we add a fatuous auxiliary task, to a network which aims to solve a semantic task such as classification or detection, the filters devoted to solving this frivolous task would not be relevant for solving the main task of concern. These filters could be pruned and pruning these would not reduce the performance on the original task. We demonstrate that this strategy is not only successful, it in fact allows for improved performance for a variety of tasks such as object classification, detection and action recognition. An interesting observation is that the task needs to be fatuous so that any semantically meaningful filters would not be relevant for solving this task. We thoroughly evaluate our proposed approach on different architectures (LeNet, VGG-16, ResNet, Faster RCNN, SSD-512, C3D, and MobileNet V2) and datasets (MNIST, CIFAR, ImageNet, GTSDB, COCO, and UCF101) and demonstrate its generalizability through extensive experiments. Moreover, our compressed models can be used at run-time without requiring any special libraries or hardware. Our model compression method reduces the number of FLOPS by an impressive factor of 6.03X and GPU memory footprint by more than 17X for VGG-16, significantly outperforming other state-of-the-art filter pruning methods. We demonstrate the usability of our approach for 3D convolutions and various vision tasks such as object classification, object detection, and action recognition.

论文关键词:Filter pruning,Model compression,Convolutional neural network,Image recognition,Deep learning

论文评审过程:Received 15 November 2019, Accepted 21 November 2019, Available online 27 November 2019, Version of Record 13 December 2019.

论文官网地址:https://doi.org/10.1016/j.imavis.2019.103857