Hierarchy
- There are lots of tasks that we plan to work on, but one slotted
for the near future is hierarchy.
- Hierarchy probably evolves in a network in many different ways
- One way is by mere association
- The dog category grows from seeing a lot of instances where
features are similar
- Simultaneously, a mammal category grows by seeing more instances
where fewer features evolve
- The question is how to learn this CA structure in a net and how
is it useful
- We might get some insight from SOMs
- Try a SOM with 25 output nodes for base level category, and one
with 4 output nodes for supercategory.
- We'd like to have a system with associative memory, so that hierarchy
actually gives you something
- I suspect I could arrange a bit string that fits say 6 possible states
(bird, mammal, zebra, lion, sparrow and cardinal) using the Hopfield
trick
- Presenting a zebra would activate zebra and mammal
- Presenting a kangaroo would activate just mammal
- You could also not present one of the input features (say birth
mode) and have the system tell you how kangaroos reproduce.
