Hierarchical Categories
- This is some work that I've submitted to the Journal of
Neural Computation, had it rejected (3 times), submitted to Connection
Science, had it rejected, and will resubmit there shortly.
- The idea is that you should be able to have subcategorisation
hierarchies in CAs.
- The network is only presented with instances of the leaf categories,
but learns both these leaf categories and the supercategory.
- It's all done with overlapping CAs.
- We named the categories cat, dog, rat and mammal. Here's a dog
- The columns represent features, Cat is 0-3-4-5-8, Dog 1-3-4-6-9,
Rat 2-3-4-7-9, and Mammal 3-4-9.
- For this to work, all four need to be stable states.
- We train it by presenting individuals, and test by presenting
individuals.
- We tested 10 different nets with 9 subordinate category items, and
allowed the net to run for 5 steps.
- 87 of the 90 are correctly categorised (based on pearson comparisons
to other members of that net).
- Similarly, presenting just the super-category led to that
state being activated.
