In ARUN-1B -- NB different from usual room

Unique chance to preview two talks due to be presented at SAB 2008 in Japan

========================

Paul Chorley presents:

Closing the Sensory-Motor Loop on Dopamine Signalled Reinforcement Learning

Paul Chorley and Anil K Seth

It has been shown recently that dopamine signalled modulation of spike timing-dependent synaptic plasticity (DA-STDP) can enable reinforcement learning of delayed stimulus-reward associations when both stimulus and reward are delivered at precisely timed intervals. Here, we test whether a similar model can support learning in an embodied context, in which timing of both sensory input and delivery of reward depend on the agent’s behaviour. We show that effective reinforcement learning is indeed possible, but only when stimuli are gated so as to occur as near-synchronous patterns of neural activity and when neuroanatomical constraints are imposed which predispose agents to exploratative behaviours. Extinction of learned responses in this model is subsequently shown to result from agent-environment interactions and not directly from any specific neural mechanism.

========================

Peter Fine presents:

Monostable controllers for adaptive behaviour

Christopher L. Buckley, Peter Fine, Seth Bullock and Ezequiel Di Paolo

Recent artificial neural networks for machine learning have exploited transient dynamics around globally stable attractors (i.e. networks possesing not more than one attractor), inspired by the properties of cortical microcolumns. Here we explore whether similarly constrained neural network controllers can be exploited for embodied, situated adaptive behaviour. We demonstrate that it is possible to evolve globally stable solutions that exhibit multiple modes of behaviour by exploiting interaction between environmental input and transient dynamics in a single basin of attraction. We present results that suggest that this globally stable regime may constitute an evolvable and dynamically rich subset of recurrent neural network configurations, especially in larger networks. We discuss the issue of scalability and the possibility that there may be alternative adaptive behaviour tasks that are more ‘attractor hungry’.