Deep Model Compression and Inference Speedup of Sum-Product Networks on Tensor Trains

Ching Yun Ko; Cong Chen; Zhuolun He; Yuke Zhang; Kim Batselier; Ngai Wong

doi:10.1109/TNNLS.2019.2928379

Deep Model Compression and Inference Speedup of Sum-Product Networks on Tensor Trains

Ching Yun Ko, Cong Chen, Zhuolun He, Yuke Zhang, Kim Batselier, Ngai Wong^*

^*Corresponding author for this work

Team Jan-Willem van Wingerden

Research output: Contribution to journal › Article › Scientific › peer-review

3 Citations (Scopus)

34 Downloads (Pure)

Abstract

Sum-product networks (SPNs) constitute an emerging class of neural networks with clear probabilistic semantics and superior inference speed over other graphical models. This brief reveals an important connection between SPNs and tensor trains (TTs), leading to a new canonical form which we call tensor SPNs (tSPNs). Specifically, we demonstrate the intimate relationship between a valid SPN and a TT. For the first time, through mapping an SPN onto a tSPN and employing specially customized optimization techniques, we demonstrate improvements up to a factor of 100 on both model compression and inference speedup for various data sets with negligible loss in accuracy.

Original language	English
Pages (from-to)	2665-2671
Journal	IEEE Transactions on Neural Networks and Learning Systems
Volume	31
Issue number	7
DOIs	https://doi.org/10.1109/TNNLS.2019.2928379
Publication status	Published - 2020

Bibliographical note

Green Open Access added to TU Delft Institutional Repository ‘You share, we take care!’ – Taverne project https://www.openaccess.nl/en/you-share-we-take-care

Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.

Keywords

Model compression
sum-product network (SP)
tensor train (TT)

Access to Document

10.1109/TNNLS.2019.2928379

08793233Final published version, 759 KB

Cite this

@article{542077fa62f14fb58a05cb381160dc91,

title = "Deep Model Compression and Inference Speedup of Sum-Product Networks on Tensor Trains",

abstract = "Sum-product networks (SPNs) constitute an emerging class of neural networks with clear probabilistic semantics and superior inference speed over other graphical models. This brief reveals an important connection between SPNs and tensor trains (TTs), leading to a new canonical form which we call tensor SPNs (tSPNs). Specifically, we demonstrate the intimate relationship between a valid SPN and a TT. For the first time, through mapping an SPN onto a tSPN and employing specially customized optimization techniques, we demonstrate improvements up to a factor of 100 on both model compression and inference speedup for various data sets with negligible loss in accuracy. ",

keywords = "Model compression, sum-product network (SP), tensor train (TT)",

author = "Ko, {Ching Yun} and Cong Chen and Zhuolun He and Yuke Zhang and Kim Batselier and Ngai Wong",

note = "Green Open Access added to TU Delft Institutional Repository {\textquoteleft}You share, we take care!{\textquoteright} – Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.",

year = "2020",

doi = "10.1109/TNNLS.2019.2928379",

language = "English",

volume = "31",

pages = "2665--2671",

journal = "IEEE Transactions on Neural Networks and Learning Systems",

issn = "2162-237X",

publisher = "IEEE Computational Intelligence Society",

number = "7",

}

TY - JOUR

T1 - Deep Model Compression and Inference Speedup of Sum-Product Networks on Tensor Trains

AU - Ko, Ching Yun

AU - Chen, Cong

AU - He, Zhuolun

AU - Zhang, Yuke

AU - Batselier, Kim

AU - Wong, Ngai

N1 - Green Open Access added to TU Delft Institutional Repository ‘You share, we take care!’ – Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public.

PY - 2020

Y1 - 2020

N2 - Sum-product networks (SPNs) constitute an emerging class of neural networks with clear probabilistic semantics and superior inference speed over other graphical models. This brief reveals an important connection between SPNs and tensor trains (TTs), leading to a new canonical form which we call tensor SPNs (tSPNs). Specifically, we demonstrate the intimate relationship between a valid SPN and a TT. For the first time, through mapping an SPN onto a tSPN and employing specially customized optimization techniques, we demonstrate improvements up to a factor of 100 on both model compression and inference speedup for various data sets with negligible loss in accuracy.

AB - Sum-product networks (SPNs) constitute an emerging class of neural networks with clear probabilistic semantics and superior inference speed over other graphical models. This brief reveals an important connection between SPNs and tensor trains (TTs), leading to a new canonical form which we call tensor SPNs (tSPNs). Specifically, we demonstrate the intimate relationship between a valid SPN and a TT. For the first time, through mapping an SPN onto a tSPN and employing specially customized optimization techniques, we demonstrate improvements up to a factor of 100 on both model compression and inference speedup for various data sets with negligible loss in accuracy.

KW - Model compression

KW - sum-product network (SP)

KW - tensor train (TT)

UR - http://www.scopus.com/inward/record.url?scp=85088037212&partnerID=8YFLogxK

U2 - 10.1109/TNNLS.2019.2928379

DO - 10.1109/TNNLS.2019.2928379

M3 - Article

C2 - 31403446

SN - 2162-237X

VL - 31

SP - 2665

EP - 2671

JO - IEEE Transactions on Neural Networks and Learning Systems

JF - IEEE Transactions on Neural Networks and Learning Systems

IS - 7

ER -

Deep Model Compression and Inference Speedup of Sum-Product Networks on Tensor Trains

Abstract

Bibliographical note

Keywords

Access to Document

Other files and links

Fingerprint

Cite this