KIP-Veröffentlichungen

Analog neuromorphic hardware is subject to fixed-pattern noise stemming from the manufacturing process and resulting in different analog behaviour in identically designed components. Calibration counteracts this mismatch by finding a set of hardware parameters that yield a desired behaviour. BrainScaleS-2 is a mixed-signal neuromorphic hardware platform emulating spiking neural networks. The current calibration framework for BrainScaleS-2 only supports single operation point calibrations, meaning that for each different calibration target, a new calibration needs to be run, which is time-consuming. Thus, the goal of this thesis is to start developing a parameter transformation model which supplies hardware parameter settings for arbitrary model parameters. As a proof of concept the transformation model is constructed and evaluated on two parameters of the leaky integrate-and-fire neuron. As these two parameters exhibit dependencies on each other’s hardware parameter, a joint transformation is developed. Even though the calibration using the transformation shows some systematic deviations, its accuracy is comparable to the fixed-point calibration leading to the conclusion that the results indicate potential for a transformation model encompassing all parameters.