Modelling optimal synaptic plasticity rules underlying associative memory.

How can intelligent systems learn? What learning rules and information coding strategies are best for allowing systems to learn that certain stimuli or actions are good or bad, even with imperfect inputs or limited, unreliable computational resources? Solutions to this engineering problem may find inspiration in biology: the humble fruit fly has solved this problem, as it can learn to associate specific odours with reward or punishment with a brain of only 100,000 neurons. This learning occurs by synaptic plasticity: changing connection strength between odour-encoding neurons and output neurons that lead to approach or avoidance behaviour. During reward learning, synapses to avoidance outputs are weakened; during punishment learning, synapses to approach outputs are weakened. Why does the fly use opposing output channels (approach and avoidance)? Why does learning weaken, rather than strengthen, synapses? We ask whether these synaptic learning rules are in some sense 'optimal': are they better than alternative rules? By what measures? Under what conditions? We will address these questions experimentally, by measuring synaptic plasticity rules for different output neurons, and computationally, by modelling how different rules perform under different conditions. The answers will inform the design of efficient learning algorithms for artificial intelligence.