General-Purpose Learning Algorithms for Spiking Neural Networks

How does the nervous system work? How does a cognitive system learn? And how is high-level human or animal learning related to changes in the nervous system? This research project will contribute to these research questions in context of research areas Computational Neuroscience, Cognitive Science and Machine Learning. We want to develop general-purpose learning algorithms for spiking neural networks. Such learning algorithms are of great importance for their potential to tie together complementary approaches towards learning on the neuronal and cognitive level and could lead to a major break-through towards a unified understanding of learning and information processing in Computational Neuroscience and Cognitive Science. The aim of Computational Neuroscience is to understand the detailed computational properties of nervous systems and build artificial neural network models that are biologically plausible, i.e. that model the function of a real neural network (in the brain) as closely as possible with an artificial neural network (on the computer). In contrast, Cognitive Science looks at (human, animal or even artificial agent) cognitive behaviour from a more global point of view and tries to draw conclusions about the underlying mechanisms of information processing in the brain. Models for such processing are often inspired by neural models, but not necessarily biologically realistic, and it is an open problem how properties of cognitive systems are grounded in properties of neural systems. On the cognitive (and also the technical) level, learning is often target-driven: a system needs to achieve a certain task, and gets feedback about how well it is doing. Based on this feedback, its behaviour is changed. Such learning also often involves inferring a priori arbitrary relations in the data given to the system. On the neural level, there are the neurons and their connections (synapses), and neuroscience has observed a number of ways in which these change when a system learns. It is however unknown how feedback on performance on the global level is broken down into localised changes to neurons and synapses on the neural level in a functional way and how known mechanisms of adaptability on this neural level conspire so that on the high-level goal-oriented learning emerges.More specifically, we want to develop learning algorithms for artificial networks of spiking neurons that make use of known neural processes of adaptability in a way such that the networks are able to learn tasks in a goal-oriented, target-driven way. Furthermore algorithms shall allow for networks to develop internal representations of a task which can be analysed and conclusions drawn from about human or animal information processing in a similar cognitive task. The project will deliver a series of learning algorithms for artificial networks of spiking neurons that are general-purpose (that is, not tied to a specific task but able to learn arbitrary input-output relationships), supervised and biologically plausible. No such algorithms exist so far. The research will have significance for the following:1. It grounds higher-level learning in low-level neural adaptability.2. The project can trigger experiments into a novel combination of learning mechanisms in the nervous system.3. It can bring forward the interpretation of the neural code through analysis of internal network dynamics in response to a learnt task.4. Models of neural systems are of interest as learning devices in their own right with a range of applications in artificial intelligence. New learning algorithms for artificial neural networks can bring forward the quest for intelligent computers.5. The research can contribute to understanding the (mal)functioning of the nervous system better, and it could consequently have a long-term impact on the medical sciences for curing neuronal disorders.

he research proposed is of theoretical and methodological nature. Its direct impact on a short time scale will be mainly on its academic beneficiaries. However the Foresight Cognitive Systems project explored on the behalf of the Government the future of research in cognitive and nervous systems and identified two main beneficiaries from this research, namely the health and information technology industries. The proposed research is part of a broad effort to understand neural processing, the brain, cognition and hence ourselves better. While much research into cognitive and neural systems is currently at a basic and fundamental level, a few first applications are reaching commercial maturity (e.g. for prosthetics and in the computer game industry), and Government expects these fields to unfold their full bearing on society and the economy in the long run (during the next 5--15 years). Consequently the long-term impact of the proposed research is seen as follows: 1. For Machine Learning, it delivers learning algorithms and models that then can be used to implement intelligent information processing systems. 2. At the same time it helps understand the brain better, potentially contributing to the Health industry's improved understanding of cognitive and nervous (mal)functioning or leading to strategies for a better use of cognitive resources. 3. It can also contribute to nanotechnology where error-tolerant distributed computation is required to build intelligent collective behaviour into a swarm of nanomachines. Use of spiking neural networks can reduce sizes of networks as well as lead to improved on-line performance. Finally, also societal interest in research into cognitive and nervous systems is high since it relates to the core of human nature and an interested public participates in the discussion of questions like Do we have a free will? or Can we read somebody else's thoughts? .