Motivation
- We really are interested in building a full-fledged AI from
simulated neurons.
- We realise this is an enormous task, and are trying to move
towards this in reasonable steps.
- We showed that a neural system using cell assemblies is Turing complete.
- We have built a games agent that viewed the environment, moved, planned,
and processed language all from simulated FLIF neurons.
- The problem was that it didn't learn much.
- In particular, learning was either one shot, or learned from
neurons directly
stimulated from the environment.
- We're also interested in reasonable neural models. We need them
to be efficient, but we also want them to be biologically
accurate.
- We're obviously short of compartmental models, but the FLIF model
has been used to model spikes.
- By extending the neural model, we get a better fit to spike data
and spontaneous activation that enables us to learn beyond the
interface.
- So, the hope is that this learning work will lead in the next few
years to games agents (or robots) that learn at many more (and maybe
all) levels.
