Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype

Chen Liu, Guillaume Emmanuel Fernand Bellec, Bernhard Vogginger, David Kappel, Johannes Partzsch, Felix Neumärker, Sebastian Höppner, Wolfgang Maass, Steve B. Furber, Robert Legenstein, Christian G. Mayr

Research output: Contribution to journalArticleResearchpeer-review

Abstract

The memory requirement of deep learning algorithms is considered incompatible with the memory restriction of energy-efficient hardware. A low memory footprint can be achieved by pruning obsolete connections or reducing the precision of connection strengths after the network has been trained. Yet, these techniques are not applicable to the case when neural networks have to be trained directly on hardware due to the hard memory constraints. Deep Rewiring (DEEP R) is a training algorithm which continuously rewires the network while preserving very sparse connectivity all along the training procedure.We apply DEEP R to a deep neural network implementation on a prototype chip of the 2nd
generation SpiNNaker system. The local memory of a single core on this chip is limited to 64 KB and a deep network architecture is trained entirely within this constraint without the use of external memory. Throughout training, the proportion of active connections is limited to 1.3%. On the handwritten digits dataset MNIST, this extremely sparse network achieves 96.6% classification accuracy at convergence. Utilizing the multi-processor feature of the SpiNNaker system, we found very good scaling in terms of computation time, per-core memory consumption, and energy constraints. When compared to a X86 CPU implementation, neural network training on the SpiNNaker 2 prototype improves
power and energy consumption by two orders of magnitude.
Original languageEnglish
Article number840
Number of pages15
JournalFrontiers in neuroscience
Volume12
Issue number840
DOIs
Publication statusPublished - 19 Nov 2018

Fingerprint

Learning
Datasets

Keywords

  • deep rewiring
  • pruning
  • sparsity
  • SpiNNaker
  • memory footprint
  • parallelism
  • energy efficient hardware

Cite this

Liu, C., Bellec, G. E. F., Vogginger, B., Kappel, D., Partzsch, J., Neumärker, F., ... Mayr, C. G. (2018). Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype. Frontiers in neuroscience, 12(840), [840]. https://doi.org/10.3389/fnins.2018.00840

Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype. / Liu, Chen; Bellec, Guillaume Emmanuel Fernand; Vogginger, Bernhard; Kappel, David; Partzsch, Johannes; Neumärker, Felix; Höppner, Sebastian; Maass, Wolfgang; Furber, Steve B.; Legenstein, Robert; Mayr, Christian G.

In: Frontiers in neuroscience, Vol. 12, No. 840, 840, 19.11.2018.

Research output: Contribution to journalArticleResearchpeer-review

Liu, C, Bellec, GEF, Vogginger, B, Kappel, D, Partzsch, J, Neumärker, F, Höppner, S, Maass, W, Furber, SB, Legenstein, R & Mayr, CG 2018, 'Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype' Frontiers in neuroscience, vol. 12, no. 840, 840. https://doi.org/10.3389/fnins.2018.00840
Liu C, Bellec GEF, Vogginger B, Kappel D, Partzsch J, Neumärker F et al. Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype. Frontiers in neuroscience. 2018 Nov 19;12(840). 840. https://doi.org/10.3389/fnins.2018.00840
Liu, Chen ; Bellec, Guillaume Emmanuel Fernand ; Vogginger, Bernhard ; Kappel, David ; Partzsch, Johannes ; Neumärker, Felix ; Höppner, Sebastian ; Maass, Wolfgang ; Furber, Steve B. ; Legenstein, Robert ; Mayr, Christian G. / Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype. In: Frontiers in neuroscience. 2018 ; Vol. 12, No. 840.
@article{8a02314c19b84c669c1755478f0ae39b,
title = "Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype",
abstract = "The memory requirement of deep learning algorithms is considered incompatible with the memory restriction of energy-efficient hardware. A low memory footprint can be achieved by pruning obsolete connections or reducing the precision of connection strengths after the network has been trained. Yet, these techniques are not applicable to the case when neural networks have to be trained directly on hardware due to the hard memory constraints. Deep Rewiring (DEEP R) is a training algorithm which continuously rewires the network while preserving very sparse connectivity all along the training procedure.We apply DEEP R to a deep neural network implementation on a prototype chip of the 2ndgeneration SpiNNaker system. The local memory of a single core on this chip is limited to 64 KB and a deep network architecture is trained entirely within this constraint without the use of external memory. Throughout training, the proportion of active connections is limited to 1.3{\%}. On the handwritten digits dataset MNIST, this extremely sparse network achieves 96.6{\%} classification accuracy at convergence. Utilizing the multi-processor feature of the SpiNNaker system, we found very good scaling in terms of computation time, per-core memory consumption, and energy constraints. When compared to a X86 CPU implementation, neural network training on the SpiNNaker 2 prototype improvespower and energy consumption by two orders of magnitude.",
keywords = "deep rewiring, pruning, sparsity, SpiNNaker, memory footprint, parallelism, energy efficient hardware",
author = "Chen Liu and Bellec, {Guillaume Emmanuel Fernand} and Bernhard Vogginger and David Kappel and Johannes Partzsch and Felix Neum{\"a}rker and Sebastian H{\"o}ppner and Wolfgang Maass and Furber, {Steve B.} and Robert Legenstein and Mayr, {Christian G.}",
year = "2018",
month = "11",
day = "19",
doi = "10.3389/fnins.2018.00840",
language = "English",
volume = "12",
journal = "Frontiers in neuroscience",
issn = "1662-4548",
publisher = "Frontiers Research Foundation",
number = "840",

}

TY - JOUR

T1 - Memory-Efficient Deep Learning on a SpiNNaker 2 Prototype

AU - Liu, Chen

AU - Bellec, Guillaume Emmanuel Fernand

AU - Vogginger, Bernhard

AU - Kappel, David

AU - Partzsch, Johannes

AU - Neumärker, Felix

AU - Höppner, Sebastian

AU - Maass, Wolfgang

AU - Furber, Steve B.

AU - Legenstein, Robert

AU - Mayr, Christian G.

PY - 2018/11/19

Y1 - 2018/11/19

N2 - The memory requirement of deep learning algorithms is considered incompatible with the memory restriction of energy-efficient hardware. A low memory footprint can be achieved by pruning obsolete connections or reducing the precision of connection strengths after the network has been trained. Yet, these techniques are not applicable to the case when neural networks have to be trained directly on hardware due to the hard memory constraints. Deep Rewiring (DEEP R) is a training algorithm which continuously rewires the network while preserving very sparse connectivity all along the training procedure.We apply DEEP R to a deep neural network implementation on a prototype chip of the 2ndgeneration SpiNNaker system. The local memory of a single core on this chip is limited to 64 KB and a deep network architecture is trained entirely within this constraint without the use of external memory. Throughout training, the proportion of active connections is limited to 1.3%. On the handwritten digits dataset MNIST, this extremely sparse network achieves 96.6% classification accuracy at convergence. Utilizing the multi-processor feature of the SpiNNaker system, we found very good scaling in terms of computation time, per-core memory consumption, and energy constraints. When compared to a X86 CPU implementation, neural network training on the SpiNNaker 2 prototype improvespower and energy consumption by two orders of magnitude.

AB - The memory requirement of deep learning algorithms is considered incompatible with the memory restriction of energy-efficient hardware. A low memory footprint can be achieved by pruning obsolete connections or reducing the precision of connection strengths after the network has been trained. Yet, these techniques are not applicable to the case when neural networks have to be trained directly on hardware due to the hard memory constraints. Deep Rewiring (DEEP R) is a training algorithm which continuously rewires the network while preserving very sparse connectivity all along the training procedure.We apply DEEP R to a deep neural network implementation on a prototype chip of the 2ndgeneration SpiNNaker system. The local memory of a single core on this chip is limited to 64 KB and a deep network architecture is trained entirely within this constraint without the use of external memory. Throughout training, the proportion of active connections is limited to 1.3%. On the handwritten digits dataset MNIST, this extremely sparse network achieves 96.6% classification accuracy at convergence. Utilizing the multi-processor feature of the SpiNNaker system, we found very good scaling in terms of computation time, per-core memory consumption, and energy constraints. When compared to a X86 CPU implementation, neural network training on the SpiNNaker 2 prototype improvespower and energy consumption by two orders of magnitude.

KW - deep rewiring

KW - pruning

KW - sparsity

KW - SpiNNaker

KW - memory footprint

KW - parallelism

KW - energy efficient hardware

U2 - 10.3389/fnins.2018.00840

DO - 10.3389/fnins.2018.00840

M3 - Article

VL - 12

JO - Frontiers in neuroscience

JF - Frontiers in neuroscience

SN - 1662-4548

IS - 840

M1 - 840

ER -