SPANNER: A Self-Repairing Spiking NeuralNetwork Hardware Architecture

Research output: Contribution to journalArticle

8 Citations (Scopus)

Abstract

Recent research has shown that a glial cell ofastrocyte underpins a self-repair mechanism in the human brainwhere spiking neurons provide direct and indirect feedbacks topre-synaptic terminals. These feedbacks modulate the synaptictransmission probability of release (PR). When synaptic faultsoccur the neuron becomes silent or near silent due to the low PR ofsynapses; whereby the PRs of remaining healthy synapses arethen increased by the indirect feedback from the astrocyte cell. Inthis paper, a novel hardware architecture of Self-rePAiringspiking Neural NEtwoRk (SPANNER) is proposed, which mimicsthis self-repairing capability in the human brain. This paperdemonstrates that the hardware can self-detect and self-repairsynaptic faults without the conventional components for the faultdetection and fault repairing. Experimental results show thatSPANNER can maintain the system performance with faultdensities of up to 40%, and more importantly SPANNER has onlya 20% performance degradation when the self-repairingarchitecture is significantly damaged at a fault density of 80%.
LanguageEnglish
Pages1287-1300
JournalIEEE Transactions on Neural Networks and Learning Systems
Volume29
Issue number4
Early online date6 Mar 2017
DOIs
Publication statusPublished - Apr 2018

Fingerprint

Neural networks
Feedback
Hardware
Neurons
Brain
Repair
Degradation
Astrocytes
Neuroglia

Keywords

  • fault tolerant computing
  • neural nets
  • SPANNER
  • astrocyte cells
  • astrocyte-neuron networks
  • fault tolerance techniques
  • fine-grained repair capability
  • self-detect faults

Cite this

@article{9f6a35db74494d52b6f039186268dee4,
title = "SPANNER: A Self-Repairing Spiking NeuralNetwork Hardware Architecture",
abstract = "Recent research has shown that a glial cell ofastrocyte underpins a self-repair mechanism in the human brainwhere spiking neurons provide direct and indirect feedbacks topre-synaptic terminals. These feedbacks modulate the synaptictransmission probability of release (PR). When synaptic faultsoccur the neuron becomes silent or near silent due to the low PR ofsynapses; whereby the PRs of remaining healthy synapses arethen increased by the indirect feedback from the astrocyte cell. Inthis paper, a novel hardware architecture of Self-rePAiringspiking Neural NEtwoRk (SPANNER) is proposed, which mimicsthis self-repairing capability in the human brain. This paperdemonstrates that the hardware can self-detect and self-repairsynaptic faults without the conventional components for the faultdetection and fault repairing. Experimental results show thatSPANNER can maintain the system performance with faultdensities of up to 40{\%}, and more importantly SPANNER has onlya 20{\%} performance degradation when the self-repairingarchitecture is significantly damaged at a fault density of 80{\%}.",
keywords = "fault tolerant computing, neural nets, SPANNER, astrocyte cells, astrocyte-neuron networks, fault tolerance techniques, fine-grained repair capability, self-detect faults",
author = "Junxiu Liu and J Harkin and LP Maguire and LJ McDaid and John Wade",
year = "2018",
month = "4",
doi = "10.1109/TNNLS.2017.2673021",
language = "English",
volume = "29",
pages = "1287--1300",
number = "4",

}

SPANNER: A Self-Repairing Spiking NeuralNetwork Hardware Architecture. / Liu, Junxiu; Harkin, J; Maguire, LP; McDaid, LJ; Wade, John.

Vol. 29, No. 4, 04.2018, p. 1287-1300.

Research output: Contribution to journalArticle

TY - JOUR

T1 - SPANNER: A Self-Repairing Spiking NeuralNetwork Hardware Architecture

AU - Liu, Junxiu

AU - Harkin, J

AU - Maguire, LP

AU - McDaid, LJ

AU - Wade, John

PY - 2018/4

Y1 - 2018/4

N2 - Recent research has shown that a glial cell ofastrocyte underpins a self-repair mechanism in the human brainwhere spiking neurons provide direct and indirect feedbacks topre-synaptic terminals. These feedbacks modulate the synaptictransmission probability of release (PR). When synaptic faultsoccur the neuron becomes silent or near silent due to the low PR ofsynapses; whereby the PRs of remaining healthy synapses arethen increased by the indirect feedback from the astrocyte cell. Inthis paper, a novel hardware architecture of Self-rePAiringspiking Neural NEtwoRk (SPANNER) is proposed, which mimicsthis self-repairing capability in the human brain. This paperdemonstrates that the hardware can self-detect and self-repairsynaptic faults without the conventional components for the faultdetection and fault repairing. Experimental results show thatSPANNER can maintain the system performance with faultdensities of up to 40%, and more importantly SPANNER has onlya 20% performance degradation when the self-repairingarchitecture is significantly damaged at a fault density of 80%.

AB - Recent research has shown that a glial cell ofastrocyte underpins a self-repair mechanism in the human brainwhere spiking neurons provide direct and indirect feedbacks topre-synaptic terminals. These feedbacks modulate the synaptictransmission probability of release (PR). When synaptic faultsoccur the neuron becomes silent or near silent due to the low PR ofsynapses; whereby the PRs of remaining healthy synapses arethen increased by the indirect feedback from the astrocyte cell. Inthis paper, a novel hardware architecture of Self-rePAiringspiking Neural NEtwoRk (SPANNER) is proposed, which mimicsthis self-repairing capability in the human brain. This paperdemonstrates that the hardware can self-detect and self-repairsynaptic faults without the conventional components for the faultdetection and fault repairing. Experimental results show thatSPANNER can maintain the system performance with faultdensities of up to 40%, and more importantly SPANNER has onlya 20% performance degradation when the self-repairingarchitecture is significantly damaged at a fault density of 80%.

KW - fault tolerant computing

KW - neural nets

KW - SPANNER

KW - astrocyte cells

KW - astrocyte-neuron networks

KW - fault tolerance techniques

KW - fine-grained repair capability

KW - self-detect faults

U2 - 10.1109/TNNLS.2017.2673021

DO - 10.1109/TNNLS.2017.2673021

M3 - Article

VL - 29

SP - 1287

EP - 1300

IS - 4

ER -