Johannes Bill: Unsupervised learning in spiking neural networks with memristor synapses

Friday, March 24, 2017 - 1:00pm to 2:00pm
Johannes Bill
Johannes Bill studied physics at Heidelberg University, Germany. He received his PhD degree in engineering at the Institute for Theoretical Computer Science at TU Graz, Austria, under the supervision of Robert Legenstein in the group of Wolfgang Maass. His PhD studies were co-funded by FACETS-ITN, a fellowship within the Marie Curie Actions program of the European Commission. Currently, he is a postdoc at the Kirchhoff-Institute for Physics, Heidelberg, in the group of Karlheinz Meier. In his research, Johannes Bill brings together computational neuroscience, machine learning theory and neuromorphic engineering. His focus is on the theory of information processing and self-organized learning in biological and artificial spiking neural networks. Specifically, he employs machine learning to understand how distributed networks without a central master or clock can adapt to their environment purely based on information that is locally available to neurons and synapses. With his studies on robust adaptation and the emergence of statistically correct learning in networks of unreliable and noisy constituents, he hopes to contribute to our understanding of perception, learning and decision making in the human brain. Besides neuroscience, Johannes Bill develops novel operation concepts for bio-inspired computing architectures, such as neuromorphic hardware, and plastic materials, such as memristors, to extend artificial information processing beyond the boundaries of von-Neumann computing.

Abstract: In an increasingly data-rich world, brain-inspired computing concepts operating neuromorphic hardware have shown great promise for processing large datasets within tight volume and power budgets. A key challenge in neuromorphic computing is the development of microscale plastic artificial synapses to extend the application of neuromorphic systems from pure processing to statistical data modeling and online learning. Memristive materials have been proposed to fulfill this need due their intrinsic electrical property of activity-dependent resistive state changes.
In this talk, I discuss a recent study that demonstrates unsupervised learning in spiking neural networks with multi-state memristor synapses. The employed titanium dioxide-based memristors exhibit weight-dependent and spike-timing dependent plasticity. This enables a memristor synapse to identify and encode in its resistive state conditional probabilities between pre- and post-synaptic neuronal activity. We demonstrate unsupervised competitive learning and classification of unlabeled noisy data in a small stochastic winner-takes-all network architecture with memristor synapses. Finally, we address some future prospects and challenges for plastic neuromorphic computing with memristors.