Proposed Problem
- We're thinking about a good experiment to get traction on
better models of CAs.
- A good next step would be to make an intermediate sized
memory model.
- There is some thing to be said about using the agent, but
the IO is tricky.
- I'm proposing a semantic net of CAs.
- Initial items would be presented derived from, for example,
word net.
- Instances would also be needed, for example Messi as a football
player.
- Associations would later be presented, based on real world
data from, for example, newspaper articles, by presenting
a small number of items simultaneously and in close temporal
sequence.
- Persistence, including STM, and synchrony could be tested by individual
presentation.
- Generalisation and usefulness of the semantic net could be tested
by priming (cat, child -> kitten) or completion (Messi plays for ?)
tasks.
- The size would have to be reasonably large, e.g. 1000 concepts and
1,000,000 neurons.
- Does that sound like a convincing (yet achievable) task?
- If that worked, would models using it be reasonable predictions
of (certain types of) memory usage?
