Learning
- The big thing that's missing from this architecture is learning.
- We (at Mdx) have done Hebbian learning, and made categorisers (e.g.
Ian and I).
- We have used Hebbian learning to implement reinforcement learning
(Roman and I).
- I've used short-term potentiation for binding (2009).
- Our learning really needs more work.
- We don't make much use of STDP, like lots of others.
- I'm really hoping to work on associative memory and learning things
properly.
- I, unlike most, use compensatory learning.
- I've just put some code together to count co-firings and then
used the result for compensatory (and Oja) learning. I think
this makes sense as LTP/D after each firing seems biologically implausible.
- I think there may be a way forward with AM, learning CAs and associations.
This can be done with a combination of compensatory learning (via
counting) and STP/D.
- I probably should be working on that right now.
