This thesis explores the intricacies of simulating large-scale spiking neural networks on digital neuromorphic platforms, emphasising the challenge of supporting the biological plausibility of implementations in resource-constrained systems. This task is hindered not only by the scale and complexity of the brain but also by current limitations in computer technologies. Nonetheless, successful endeavours could enhance our understanding of neural mechanisms, neurological disorders, and their implications for emerging technologies. Therefore, this study proposes implementing novel simulation strategies on digital hard- ware, focusing on scalable, efficient solutions and offering a pathway to more sophisticated and accessible simulations of brain activity. The proposed implementation includes a low-precision floating-point data type for biological neural network simulations, challenging the notion that imprecise computations cannot effectively capture network dynamics and statistics. The thesis also develops neural models and procedures that optimise memory access, significantly enhancing the efficiency of simulations involving synaptic plasticity. Furthermore, a novel implementation of synaptic plasticity that leverages stochastic rounding in a low-precision context is studied. The work herein presented utilises simulation and analytical tools to analyse the pro- posed methods comprehensively. Moreover, results are substantiated by comparisons with theoretical and experimental insights from the computational neuroscience literature. Overall, the findings not only validate the proposed methods against benchmarks but also open new avenues for research in neuromorphic engineering and neuroscience.
Memory-efficient neuromorphic neurons and synapses for simulating large-scale spiking neural networks
De Abreu Urbizagastegui, P. (Author). 2024
Western Sydney University thesis: Doctoral thesis