Error Resilience Analysis for Systematically Employing Approximate Computing in Convolutional Neural Networks
Muhammad Abdullah Hanif1,a , Rehan Hafiz2 and Muhammad Shafique1,b
1Institute of Computer Engineering, Vienna University of Techoology, Vienna, Austria
amuhammad.hanif@tuwien.ac.at
bmuhammad.shafique@tuwien.ac.at
2Department of Electrical Engineering, Information Techoology University, Lahore, Pakistan
rehan.hafiz@itu.edu.pk
ABSTRACT
Approximate computing is an emerging paradigm for error n:silient applications as it leverages accuracy loss for improving power, energy, area, and/or performance of an application. The spectrum of error resilient applications includes the domains of Image and video processing, Artificial Intelligence (AI) and Machine Learning (ML), data IIJiaiytia,and other Reeogo!Uoo, Mlnlng, and Synthesis (RMS) applleatioos.In this work, we address one of the most challenging question, i.e., bow to systematically employ approximate computing in Convolution Neural Networks (CNNs), which are one of the most compute-intensive and the pivotal part of AI. Towards this, we prop01e a methodology to systematically analyze error resilience of deep CNNs and identify parameters that can be exploited for impro'ring performance/efficiency of these networks for inference purposes. We also present a case study for significance-driven claJsification of filters for different convolutional layers, and propose to prune those having the least sigDilleaooe, and thereby eosbling accuracy vs. efilciency tradeollli by exploiting their resilience characteristics in a systematic way.
