The Theory of Darwinian Neurodynamics (1)
- Speaker: Dr Chrisantha Fernando (EECS, Queen Mary University of London)
- Host: Steve Furber
- 25th October 2012 at 15:00 in Lecture Theatre 1.3, Kilburn Building
We propose that there are informational replicators in the brain, akin to DNA but quite different in implementation. Response characteristics, e.g. orientation selectivity in visual cortex, are known to be copied from neuron to neuron. Through STDP, small neuronal circuits can undertake causal inference on other circuits to reconstruct the topology of those circuits based on observations of their spontaneous activity. Patterns of synaptic connectivity are replicating units of evolution in the brain. How does this map to the cognitive architecture level? The space of predictions is unlimited; brains do sparse search in this model space. We've shown that Darwinian dynamics is efficient for sparse search compared to algorithms that lack information transfer between adaptive units, in realistic adaptive landscapes. Thus, we propose that Darwinian dynamics in the brain implements approximate Bayesian inference, is capable of search in the space of physical symbol systems, search for syntactic rules, and search over structured representations in human insight problems.