Enhancing Neuromorphic Computing with Advanced Spiking Neural Network Architectures

Abstract

This dissertation proposes ways to address current limitations of neuromorphic computing to create energy-efficient and adaptable systems for AI applications. It does so by designing novel spiking neural networks architectures that improve their performance. Specifically, the two proposed architectures address the issues of training complexity, hyperparameter selection, computational flexibility, and scarcity of neuromorphic training data. The first architecture uses auxiliary learning to improve training performance and data usage, while the second architecture leverages neuromodulation capability of spiking neurons to improve multitasking classification performance. The proposed architectures are tested on Intel's Loihi2 neuromorphic chip using several neuromorphic datasets, such as NMIST, DVSCIFAR10, and DVS128-Gesture. The presented results demonstrate potential of the proposed architectures but also reveal some of their limitations which are proposed as future research.