Other Models and CAs
- Hopfield Nets and SOMs can inform CA development
- Can other models inform CAs
- Adaptive Resonance Theory enables you to form tight or
broad categories (Hamming distance) depending on a variance
parameter
- This may be a useful concept for CA formation
- CAs may have a solution for Plasticity Stability dilemma
-
- With long term dynamics, stable states frequently
persist
- This is because neurons in a CA occasionally activate
each other and allow the Hebbian rule to increase
the connection strength
- We've been calling this bicycle riding and spontaneous
activation is enough to support this
- Still, these neurons can participate in other CAs
- Consequently, a network can learn new concepts, but
retain old ones
- Of course there are limits to the amount of CAs a net
can hold
- We've not thoroughly explored this
- Wickelgren states more than N states can be held
- We've also not done much exploration of input to the system
- We simply turn nodes on
- Layering, existing vision and other modality work may help here
