Open Peer Review. Open Publishing. Open Access. Open Discussion. Open Directory. Open Recommendations. Open API. Open Source.
Accelerating Convolutional Neural Networks using Iterative Two-Pass Decomposition
Nov 03, 2017 (modified: Nov 03, 2017)ICLR 2018 Conference Blind Submissionreaders: everyoneShow Bibtex
Abstract:We present the iterative two-pass decomposition flow to accelerate existing convolutional neural networks (CNNs). The proposed rank selection algorithm can effectively determine the proper ranks of the target convolutional layers for the low rank approximation. Our two-pass CP-decomposition helps prevent from the instability problem. The iterative flow makes the decomposition of the deeper networks systematic. The experiment results shows that VGG16 can be accelerated with a 6.2x measured speedup while the accuracy drop remains only 1.2%.
TL;DR:We present the iterative two-pass CP decomposition flow to effectively accelerate existing convolutional neural networks (CNNs).