Ten thousand times faster: Classifying multidimensional data on a spiking neuromorphic hardware system.

Discrimination of sensory inputs is a computational task that biological neuronal systems perform very efficiently. Assessing the principles in those systems is a promising approach to develop technical solutions for many problems, such as data classification. A particular problem here is to train a classifier in a supervised fashion to discriminate classes in multidimensional data. We implemented a network of spiking neurons that solves this task using a neuromorphic hardware system, that is, analog neuronal circuits on a silicon substrate. This system enables us to do high-performance computation in a biologically inspired way, with spiking neurons as computational units. In this contribution, we illustrate solutions to technical challenges that occur when implementing a classifier on neuromorphic hardware.The network topology of the insect olfactory system provides a well suited template for a neuronal architecture processing multidimensional data. In our classifier network, the value of each dimension of a data vector determines the rate of a stochastically generated spike train. The spike trains are fed into non-overlapping populations of neurons. Those populations project onto an association layer with winner-take-all properties representing the output of the classifier. During classifier training, the weights in this projection are adjusted according to a firing-rate based learning rule.The values in multidimensional data sets are typically real numbers, but neuronal firing rates are restricted to values between zero and some maximal value. Hence, the data must be transformed into a positive, bounded representation. We achieved this by representing each data point as a vector of distances to a number of points in data space (“virtual receptors” [1]). The representation by virtual receptors inevitably introduces correlation between input dimensions. We reduced this correlation using lateral inhibition in the first neuronal layer, leading to a significant increase in classifier performance. We found that decorrelation was most efficient when we scaled the inhibitory weights according to the correlation between the connected populations.We ran our classifier network on a neuromorphic hardware system that runs at ten thousand times the speed of biological neurons, thus suited for high performance computing [2]. However, the considerable variance of rate-response sensitivity across hardware neurons decreased classification performance. We therefore developed a calibration routine to counteract the neuronal variance. [1] Schmuker, M. and Schneider, G. (2007). Processing and classification of chemical data inspired by insect olfaction. Proc. Natl. Acad. Sci. U S A 104, 20285-20289.[2] Bruderle, D., Bill, J., Kaplan, B., Kremkow, J., Meier, K., Muller, E. and Schemmel, J. (2010). Simulator-like exploration of cortical network architectures with a mixed-signal VLSi system. In Proc. of IEEE Intern. Symp. on Circuits and Systems (ISCAS), 2784–8787.