Real World Tasks
- We've also done some practical software tasks.
- CAs are categorisers, so we used it on one of the
tasks in the U. Cal. Irvine database: the congressional
voting task.
- Given a list of the voting behaviours of US congressmen on 8
bills, categorise them by party.
- We used a 10 fold analysis and got 90% of them correct.
- Another task was Information Retrieval.
- Train the net on documents by having a neuron per word.
- Input was the documents, so the net learned a type of
coocurence matrix.
- Test by specifying a query of words.
- Retrieve by running the net and doing a Pearson's correlation
to find the nearest document(s).
- This performed in the range of existing IR systems on a
standard benchmark.
