Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training

This paper reports how to train a Deep Belief Network (DBN) using only 8-bit fixed-point parameters. We propose a dynamic-point stochastic rounding algorithm that provides enhanced results compared to the existing stochastic rounding. We show that by using a variable scaling factor, the fixed-point...

Full description

Main Authors: Essam, M., Tang, T.B., Ho, E.T.W., Chen, H.
Format: Article
Institution: Universiti Teknologi Petronas
Record Id / ISBN-0: utp-eprints.20033 /
Published: IEEE Computer Society 2017
Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85028585216&doi=10.1109%2fNER.2017.8008430&partnerID=40&md5=3d565633d6b0ee148b8349a9b15f1d76
http://eprints.utp.edu.my/20033/
Tags: Add Tag
No Tags, Be the first to tag this record!
id utp-eprints.20033
recordtype eprints
spelling utp-eprints.200332018-04-22T14:38:14Z Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training Essam, M. Tang, T.B. Ho, E.T.W. Chen, H. This paper reports how to train a Deep Belief Network (DBN) using only 8-bit fixed-point parameters. We propose a dynamic-point stochastic rounding algorithm that provides enhanced results compared to the existing stochastic rounding. We show that by using a variable scaling factor, the fixed-point parameter updates are enhanced. To be more hardware amenable, the use of common scaling factor at each layer of DBN is further proposed. Using publicly available MNIST database, we show that the proposed algorithm can train a 3-layer DBN with an average accuracy of 98.49, with a drop of 0.08 from the double floating-point average accuracy. © 2017 IEEE. IEEE Computer Society 2017 Article PeerReviewed https://www.scopus.com/inward/record.uri?eid=2-s2.0-85028585216&doi=10.1109%2fNER.2017.8008430&partnerID=40&md5=3d565633d6b0ee148b8349a9b15f1d76 Essam, M. and Tang, T.B. and Ho, E.T.W. and Chen, H. (2017) Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training. International IEEE/EMBS Conference on Neural Engineering, NER . pp. 629-632. http://eprints.utp.edu.my/20033/
institution Universiti Teknologi Petronas
collection UTP Institutional Repository
description This paper reports how to train a Deep Belief Network (DBN) using only 8-bit fixed-point parameters. We propose a dynamic-point stochastic rounding algorithm that provides enhanced results compared to the existing stochastic rounding. We show that by using a variable scaling factor, the fixed-point parameter updates are enhanced. To be more hardware amenable, the use of common scaling factor at each layer of DBN is further proposed. Using publicly available MNIST database, we show that the proposed algorithm can train a 3-layer DBN with an average accuracy of 98.49, with a drop of 0.08 from the double floating-point average accuracy. © 2017 IEEE.
format Article
author Essam, M.
Tang, T.B.
Ho, E.T.W.
Chen, H.
spellingShingle Essam, M.
Tang, T.B.
Ho, E.T.W.
Chen, H.
Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training
author_sort Essam, M.
title Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training
title_short Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training
title_full Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training
title_fullStr Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training
title_full_unstemmed Dynamic point stochastic rounding algorithm for limited precision arithmetic in Deep Belief Network training
title_sort dynamic point stochastic rounding algorithm for limited precision arithmetic in deep belief network training
publisher IEEE Computer Society
publishDate 2017
url https://www.scopus.com/inward/record.uri?eid=2-s2.0-85028585216&doi=10.1109%2fNER.2017.8008430&partnerID=40&md5=3d565633d6b0ee148b8349a9b15f1d76
http://eprints.utp.edu.my/20033/
_version_ 1741196305241210880
score 11.62408

Similar Items