This project demonstrates multivariate classification using neuromorphic hardware, in this case the SpiNNaker hardware platform [2]. The classifier neural model design is based on that published in [1,3,4] which is a bio-inspired design based on the insect olfactory system. Reference [1] provides the fullest details of how the classifier is implemented on the SpiNNaker platform.

The core requirements to run these programs is Python 2.7 and a SpiNNaker board with the accompanying software suite made available from Manchester University (see https://github.com/SpiNNakerManchester); The work is demonstrated here in a series of Python/PynNN/sPyNNaker experiments of increasing complexity/difficulty.

EXPERIMENT 1 This experiment demonstrates a basic ability to use synapse plasticity to learn to classify ("recognise") 10 digits taken from the standard MNIST dataset. This experiment requires only the SpyNNaker software to be functional and any attached SpiNNaker board. For those unable to run the program themselves, screenshot(s) of the relevant raster plots are made available in the "screenshots" subdirectory.

SPINN5- LARGE MODEL - LIVE SPIKING This experiment demonstrates classification of MNIST using a much larger model (tested up to 30K neurons) running on the 48-chip SPiNN5 board. To achieve the throughput of input data for this large dataset, rather than spike source files, this experiment uses a C/C++ front end program to send more than 500 channels of live spikes to the board via UDP. The C program was developed and tested on Ubuntu Linux using the g++ compiler.

ACKNOWLEDGEMENTS This work is supported by a grant from the Human Brain Project (HBP). The SpiNNaker hardware platform is developed at Manchester University who supplied the board and software suite used in this work.

CITING THIS WORK To cite the work in this project please contact one of the Github collaborators based at the University of Sussex.

REFERENCES

[1] A. Diamond, T. Nowotny and M. Schmuker, "Comparing neuromorphic solutions in action: implementing a bio-inspired solution to a benchmark classification task on three parallel-computing platforms" in Frontiers in Neuromorphic Engineering, 2016, Vol 9, No. 00491 doi:10.3389/fnins.2015.00491

[2] M. M. Khan, D. R. Lester, L. A. Plana, A. Rast, X. Jin, E. Painkras, and S. B. Furber, “SpiNNaker: Mapping neural networks onto a massively- parallel chip multiprocessor,” in Proceedings of the International Joint Conference on Neural Networks, 2008, pp. 2849–2856

[3] M. Schmuker, T. Pfeil, and M. Nawrot, “A neuromorphic network for generic multivariate data classification,” Proc. Natl. Acad. Sci., pp. 1–6, Jan. 2014.

[4] A. Diamond, M. Schmuker, A.Z. Berna, S.Trowell and T. Nowotny "Classifying chemical sensor data using GPU-accelerated bio-mimetic neuronal networks based on the insect olfactory system", BMC Neuroscience 2014, 15(Suppl 1):P77
doi:10.1186/1471-2202-15-S1-P77