Acceleration of Convolutional Networks Using Nanoscale Memristive Devices

Shruti R. Kulkarni, Anakha V. Babu, Bipin Rajendran

Research output: Chapter in Book/Report/Conference proceedingConference contribution

1 Scopus citations

Abstract

We discuss a convolutional neural network for handwritten digit classification and its hardware acceleration as an inference engine using nanoscale memristive devices in the spike domain. We study the impact of device programming variability on the spiking neural network’s (SNN) inference accuracy and benchmark its performance with an equivalent artificial neural network (ANN). We demonstrate optimization strategies to implement these networks with memristive devices with an on-off ratio as low as 10 and only 32 levels of resolution. Further, close to baseline accuracies can be maintained for the networks even if such memristive devices are used to duplicate the pre-determined kernel weights to enable parallel execution of the convolution operation.

Original languageEnglish (US)
Title of host publicationEngineering Applications of Neural Networks - 19th International Conference, EANN 2018, Proceedings
EditorsElias Pimenidis, Chrisina Jayne
PublisherSpringer Verlag
Pages240-251
Number of pages12
ISBN (Print)9783319982038
DOIs
StatePublished - Jan 1 2018
Event19th International Conference on Engineering Applications of Neural Networks, EANN 2018 - Bristol, United Kingdom
Duration: Sep 3 2018Sep 5 2018

Publication series

NameCommunications in Computer and Information Science
Volume893
ISSN (Print)1865-0929

Other

Other19th International Conference on Engineering Applications of Neural Networks, EANN 2018
CountryUnited Kingdom
CityBristol
Period9/3/189/5/18

All Science Journal Classification (ASJC) codes

  • Computer Science(all)
  • Mathematics(all)

Keywords

  • Artificial neural networks
  • Memristors
  • Non-volatile memory devices
  • Programming variability
  • Spiking neural networks

Fingerprint Dive into the research topics of 'Acceleration of Convolutional Networks Using Nanoscale Memristive Devices'. Together they form a unique fingerprint.

Cite this