Neural networks-on-chip for hybrid bio-electronic systems

Abstract

By modelling the brains computation we can further our understanding of its function and develop novel treatments for neurological disorders. The brain is incredibly powerful and energy e cient, but its computation does not t well with the traditional computer architecture developed over the previous 70 years. Therefore, there is growing research focus in developing alternative computing technologies to enhance our neural modelling capability, with the expectation that the technology in itself will also bene t from increased awareness of neural computational paradigms. This thesis focuses upon developing a methodology to study the design of neural computing systems, with an emphasis on studying systems suitable for biomedical experiments. The methodology allows for the design to be optimized according to the application. For example, di erent case studies highlight how to reduce energy consumption, reduce silicon area, or to increase network throughput. High performance processing cores are presented for both Hodgkin-Huxley and Izhikevich neurons incorporating novel design features. Further, a complete energy/area model for a neural-network-on-chip is derived, which is used in two exemplar case-studies: a cortical neural circuit to benchmark typical system performance, illustrating how a 65,000 neuron network could be processed in real-time within a 100mW power budget; and a scalable highperformance processing platform for a cerebellar neural prosthesis. From these case-studies, the contribution of network granularity towards optimal neural-network-on-chip performance is explored.