Event Abstract Back to Event A spiking classifier for nonlinear problems implemented on a neuromorphic hardware system Michael Schmuker1, 2*, Sven Schrader3, Thomas Pfeil3 and Martin P. Nawrot1, 2 1 Freie Universität Berlin, Germany 2 Bernstein Center for Computational Neuroscience Berlin, Germany 3 Universität Heidelberg, Kirchhoff Institute for Physics, Germany The question how neuronal systems process sensory information is as crucial for neuroscience as it is for bio-inspired technical applications. Classification of multidimensional data is a common problem in signal and data analysis. The architecture of the olfactory system maps particularly well onto this problem [1]. Here, we present an olfaction-inspired spiking classifier network that achieves the performance of a Naive Bayes classifier on a benchmark data set, and demonstrate its ability to solve nonlinear classification problems. Moreover, we show an implementation of this network on a neuromorphic hardware system comprising 192 spiking analog hardware neurons embedded in digital control circuitry. Neuronal computations in this system operate at a speedup factor of 10^4 compared to biological real time, enabling high-thoughput neurocomputing. The neuromorphic implementation is subject to inherent device mismatch affecting analog components, e.g. neuron parameters, which leads to a decrease of classification performance. As a solution we present an algorithm for self-calibration that takes into account the specific architecture of our network in order to maximally exploit the capabilities of the hardware system. With this algorithm we were able to increase classifier performance on the hardware to a level comparable with a Naive Bayes classifier. As neuronal variation is not only a feature of neuromorphic hardware but also a hallmark of neurons in biological systems, our results provide insight into the biological relevance of neural coding hypotheses thought to operate on a neuronal substrate with heterogeneous sensitivity.