KIP-Veröffentlichungen

Spiking neural networks can solve complex tasks in an event-based processing strategy, inspired by the brain. One special kind of neuron model, the AdEx model, allows to reproduce several types of firing patterns, which have been found in biological neurons and may be of functional importance. In this thesis we characterize the analog neuron circuit implementation of this model within the full-custom HICANN ASIC. As the central unit of the BrainScaleS accelerated neuromorphic computing platform, it provides a tool to emulate large neural networks in short time and helps to better understand the brain. Characterization of the neuron circuits leads to calibration of each sub-circuit, translating the desired AdEx model parameters to their corresponding HICANN parameters for each individual neuron. Device mismatch in VLSI manufacturing leads to expected variation from design parameters. These variations can be counteracted by adjustable parameters within the circuits. A wafer-scale BrainScaleS system contains over 1.9·10^5 neuron circuits with millions of parameters. Due to the large scale of the system, methods need to be fully automated in a robust way. Characterizations presented in this work are performed from transistor level simulation to wafer-scale hardware measurements. Our commissioning and calibration efforts are enabling neural network experiments, including complex firing patterns that are computationally expensive when implemented in traditional numerical simulations.