Neuromorphic architectures with electronic synapses

S Burc Eryilmaz1, Siddharth Joshi2, Emre Neftci3, Weier Wan1, Gert Cauwenberghs2, H.-S. Wong1
1Stanford University, 2University of California, San Diego, 3Department of Cognitive Sciences, University of California Irvine


This paper gives an overview of recent progress on 1) online learning algorithms with spiking neurons 2) neuromorphic platforms that efficiently run these algorithms with a focus on implementation using analog-non-volatile memory (aNVM) as electronic synapses. Design considerations and challenges for using aNVM synapses such as requirements for device variability, multilevel states, programming energy, array-level connectivity, wire energy, fan-in/fan-out, and IR drop are presented. Future research directions and integration challenges are summarized. Algorithms based on spiking neural networks are promising for energy efficient real-time learning, but cycle-to-cycle device variations can significantly impact learning performance. Our analysis suggests that wires are increasingly important for energy considerations, especially for large systems.