Modelling Neural Dynamics On Neuromorphic Hardware

Abstract

The brain, composed of over 86 billion neurons with diverse functions, morphology, connective properties, and communication methods, remains the most intricate organ in the human body. Simulation of brain and neural activity has seen significant progress through Spiking Neural Networks (SNNs), featuring individual mathematical models of neurons. Neuromorphic systems aim to provide accelerated low-power simulation of SNNs but primarily rely on simplified and efficient neuron models like the Leaky Integrate-and-Fire (LIF) model. Meanwhile, neuroscientists have developed biologically plausible neuron models that offer enhanced computational capabilities compared to LIF models, capturing complex neural dynamics. Unfortunately, the adoption of these biologically inspired models in neuromorphic computing has been limited due to their computational costs, where energy-efficiency stands as a key requirement. This work addresses the need for bridging the gap between neuroscience and neuromorphic computing by exploring modelling techniques and implementation strategies to make complex, biologically accurate neuron models feasible for neuromorphic hardware, and testing their efficacy within SNNs. Work centres on the Hodgkin-Huxley (HH) neuron, representing a single compartment, and a two-compartment neuron model that incorporates dendritic computation. To evaluate their efficacy, models are implemented on SpiNNaker and SpiNNaker2 prototype neuromorphic systems, benchmarking their numerical accuracy against reference models within the NEURON simulation environment. Remarkably, both the fixed- and floating-point SpiNNaker implementations displayed excellent agreement with the reference models. While the complexity of the models increased computation times compared to LIF models, the HH neuron only exhibited an 8x wall clock time increase (11x for the two-compartment model), showcasing the potential of hardware accelerators in the next-generation neuromorphic system to optimise complex neuron model implementation. Integration and training of the two-compartment neuron in SNNs is also explored in this work. The two-compartment model is trained to detect a diverse range of patterns, demonstrating a decrease in the overall energy consumption compared with SNNs of LIF neurons completing the same tasks. By showcasing the successful implementation of biologically accurate neuron models on neuromorphic hardware, this work offers a pathway for more sophisticated and powerful neuromorphic systems capable of simulating brain-like activities more efficiently. The incorporation of complex neuron models into neuromorphic systems enables research into neural behaviours not captured by simplified models. This, in turn, fosters a more profound understanding of brain mechanisms and provides more realistic representations of neural processes. The proposed neuromorphic approach, with hardware accelerators optimised for SNN simulations, offers a promising solution to address the limitations of High-Performance Computing (HPC) when simulating complex networks. Ultimately, this research contributes to advancing understanding of brain function, improving AI algorithms, and showcases the potential of hardware accelerators in next-generation neuromorphic systems.

Date of Award	11 Feb 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Oliver Rhodes (Main Supervisor) & James Garside (Co Supervisor)