Ideas About Version 1
- Assuming, whacking together the static AM and rules is a good
way forward for a proto neuro cognitive architecture, where next?
- I think I really need to focus on proper CAs as opposed to
binary CAs.
- In particular, I think these need to be learned, so I think the
AM stuff is the big issue.
- We (people developing the architecture) could also readily
integrate other systems like spatial reasoning or vision.
(They're still neural, and they probably still use CAs.)
- Even better, multiple versions of the sub systems could
be developed so that they could be compared.
- Attention
- We should also be able to learn rules.
- We might be able to create new ones (via learning FSAs Maass).
- We could also use RL to select between rules.
- There's probably something else we could do with specialising and
generalising rules (particularly if the AM worked).
- I'm really interested in using that Soar problem space
mechanism.
- There are also specialised subsystems in the brain for, among other
things, spatial memory, vision, and speech recognition. My hope
is that, if you do them in spiking neurons, at least initially,
these modules should just drop into the architecture.
- The modules have a simple communication mechanism: spikes.
- Eventually, the systems are going to have to learn together, like
a real biological neural agent does.
