Neuroscientific Modeling with a Mixed-Signal VLSI Hardware System

Abstract

Modeling networks of spiking neurons is a common scientific method that helps to understand how biological neural systems represent, process and store information. But the simulation of large-scale models on machines based on the Turing paradigm is subject to performance limitations, since it suffers from an intrinsic discrepancy to the massive parallelism of neural processing in the brain. Following an alternative approach, neuromorphic engineering implements the structure and function of biological neural systems in analog or analog-digital VLSI devices. Neuron and synapse circuits represent physical models that evolve in parallel and in continuous time. Therefore, neuromorphic systems can overcome limitations of pure software approaches in terms of speed and scalability. Recent developments aim at the realization of large-scale, massively accelerated and highly configurable neuromorphic architectures. This thesis presents a novel methodological framework that renders possible the beneficial utilization of such devices as neuroscientific modeling tools. In a comprehensive study, it describes, tests and characterizes an existing prototype in detail. It presents policies for the biological interpretation of the hardware output and techniques for the calibration of the chip. The thesis introduces a dedicated software framework that implements these methods and integrates the hardware interface into a simulator-independent modeling language, which is also supported by various established software simulators. This allows to port experiment descriptions between hardware and software simulators, to compare generated output data and consequently to verify the hardware model. The functionality of the translation methods, the calibration techniques and the verification framework are shown in various experiments both on the single cell and on the network level.