Overlapping and Orthogonal CAs
One thing we should note is that these simulations have used (largely) orthogonal CAs. That is, each CA has its neurons and no others.
It seems likely that CAs are really overlapping; they share neurons.
We've used overlapping neurons before to store hierarchical categories.
The statistical mechanics stuff (e.g. Hopfield) and some graph theoretical work (Valiant 2005) uses overlapping coding.
In the above simulations, CAs are associated by synapses between member neurons.
With overlapping CAs, sharing neurons is a form of association.
However, we don't really have a good way of assigning neurons to CAs.
Hopefully, the CABot work will enable CAs and associations to emerge from interaction with the environment.
