학술논문
ReRAM-Based NeoHebbian Synapses for Faster Training-Time-to-Accuracy Neuromorphic Hardware
Document Type
Conference
Author
Source
2023 International Electron Devices Meeting (IEDM) Electron Devices Meeting (IEDM), 2023 International. :1-4 Dec, 2023
Subject
Language
ISSN
2156-017X
Abstract
NeoHebbian artificial synapses based on ReRAM devices, enabling scalable online e-Prop learning algorithms, were proposed and experimentally demonstrated for the first time. Such synapses feature two state variables - a neuron coupling weight and "eligibility trace" for updating the weight. To implement these features, the first proposed design utilizes a pair of ReRAM devices, each with unique conductance range, and connects them in series to form a voltage divider for controlling a weight update, and in parallel for the remaining operations. The second, a denser design utilizes heater to modulate ReRAM weight update. Benchmark simulation results have shown that the proposed devices could dramatically reduce training time and energy consumption for temporal data modeling applications.