Sequences
- Hina Ghalib has been working on her PhD and will hopefully finish
in early 2009.
- She won the best student paper award at IJCNN for our model of
sequences of CAs.
- Roughly there is a CA for each possible item.
- These base items then can be combined to form pairs,
triples and so forth.
- The IJCNN paper didn't involve learning, but Hina has now developed
a system that does.
- It has some topological plausibility issues, but it
functions quite well (~95% for length 5).
- It does one shot learning.
- The memories persist for quite some time. (It's not a
short term rehearsal mechanism like Deco and Rolls.)
- I don't think this is the only way to learn sequences, but
it's nice to have one mechanism that works.
- I think her future work will involve simulating psychological
data, including chunking.
- However, there seems to be a lot that could be done on finding
varieties of sequence learning mechanisms.
