Peter O’Connor, Efstratios Gavves, Max Welling Temporally Efficient Deep Learning with Spikes May 2017 Summary: We devise a scheme for implementing an approximation to backpropagation when input data is temporally redundant. We show that this results in leaky integrate-and-fire neurons, and that the resulting learning rule is equivalent to a form of STDP. We demonstrate that we can use this to train a deep network successfully.
Peter O’Connor, Max Welling
Sigma Delta Quantized Networks
November 2016, accepted to ICLR 2017. https://arxiv.org/abs/1611.02024
Summary: We devise a scheme wheren neurons in a network communicate their quantized changes in activation, rather than their activation itself, allowing these networks to run more efficiently on temporally redundant data.
Peter O’Connor, Max Welling
Deep Spiking Networks
February 2016. Available in pre-print at: http://arxiv.org/pdf/1602.08323v1.pdf
Summary: We implement backpropagation with event-based neurons.
Peter O’Connor, D Neil, S Liu, T Delbruck, M Pfeiffer
Real-Time Classification and Sensor Fusion with a Spiking Deep Belief Network
Frontiers in Neuroscience. Vol 7, 2013. http://journal.frontiersin.org/Journal/10.3389/fnins.2013.00178
Summary: We train a Deep Belief network to Associate digits with audio tones, and map the trained network onto an event-based network of Leaky Integrate and Fire neurons. This was an early demonstration that deep networks can be implemented with spiking neurons.
R.C. Sotero, A. Bortel, R. Martinez-Cancino, S. Neupane, P. O’Connor, F. Carbonell, A. Shmuel. Anatomically-constrained effective connectivity among layers in a cortical column modeled and estimated from local field potentials. Journal of Integrated Neuroscience, 9 (2010), pp. 355-379. Available at: http://www.ncbi.nlm.nih.gov/pubmed/21213410