| PhD Seminar

Name of the Speaker: Mr. Shubham R Pande (EE18D704)
Guide: Prof. Anjan Chakravorthy
Co-Guide: Dr. Bhaswar Chakrabarti
Venue: ESB-244 (Seminar Hall)
Date/Time: 9th May 2024 (Thursday), 2:00 PM
Title: NeoHebbian Synapses for Accelerating Online Training of Neuromorphic Hardware

Abstract

We propose and experimentally validate a neoHebbian artificial synapse based on ReRAM devices to enable the hardware implementation of a scalable online E-Prop learning algorithm. This synapse features two state variables - a neuron coupling weight and an "eligibility trace" for updating the weight. We propose a heater-integrated ReRAM synapse to implement these features. The neuron coupling weight is represented by ReRAM conductance, whereas the eligibility trace is encoded in the form of local temperature at ReRAM. Benchmark simulations, taking into account various device and system-level non-idealities, have shown that the proposed synapse has the potential to significantly reduce training time and energy on temporal modeling tasks.