Fault Impact Map for Memristive Crossbar Neural Networks

Research output: Contribution to book/conference proceedings/anthology/reportConference contributionContributedpeer-review

Abstract

In this paper, we propose the Fault Impact Map (FIM) as a neural network post-training method measuring the accuracy impact of faults or stuck devices in memristive crossbar computing. Memristive crossbars enable highly energy-efficient neural network computation, but stuck-at-faults (SAF) can affect the weight representation and significantly reduce network performance. Recent studies consider sensitivity-based or randomly injected SAF and propose various mapping or training methods to improve the network performance. However, the not unlikely worst case of SAF is not taken into account, which leads to an overestimation of the robustness of the network. We introduce a novel brute-force FIM algorithm to detect the SAF worst-case scenario of a trained memristive crossbar neural network and determine the weights that are most important for maintaining its test accuracy. The FIM algorithm measures the accuracy drop caused by successively faulty weights or biases leading to a drastic accuracy drop. For an MNIST classification performed on a quantized 2-layer memristive crossbar neural network, we compare the FIM with state-of-the-art methods like gradient-based sensitivity analysis and random fault injection and discuss the prediction of the network vulnerability to SAF. Our test results show, that SAF of less than 1% of the neural network parameters are sufficient to shrink the test accuracy down to the guess line. Finally, we discuss the computationally effort of the FIM as well as its potential application to increase the network robustness.

Details

Original languageEnglish
Title of host publication2024 13th International Conference on Modern Circuits and Systems Technologies (MOCAST)
PublisherIEEE
Pages1-4
Number of pages4
ISBN (electronic)979-8-3503-8542-7
ISBN (print)979-8-3503-8543-4
Publication statusPublished - 28 Jun 2024
Peer-reviewedYes

Conference

Title13th International Conference on Modern Circuits and Systems Technologies
Abbreviated titleMOCAST 2024
Conference number13
Duration26 - 28 June 2024
Website
Degree of recognitionInternational event
LocationTechnical University of Sofia
CitySofia
CountryBulgaria

External IDs

ORCID /0000-0001-7436-0103/work/165877290
ORCID /0000-0003-3259-4571/work/165877922
Scopus 85202444428

Keywords

Keywords

  • Fault Impact Map, Memristive crossbar neural network, stuck-at-faults