Old Categorisation Results
- So, we played around with some topologies and came up with
a four net topology.
- The input (base) subnet represents the input vectors, and the output
net on the right represents the categories.
- Connections are: input->som->hidden->output; som, hidden and output are
all self connected; output->hidden-> som.
- The learning is post-compensatory from input and all the others are
pre-compensatory.
- Here's part of a neural simulation that learns irises. It
got about 93%, which is a reasonable score on this task.
- The picture shows the net during training.
- Blue circles are firing neurons, and the white squares are
neurons that don't fire in this cycle.
- One instance of an iris is presented in the base net, and
its told its category three (both by externally activating
those neurons).
- The firing in the middle two subnets is via spread of activation
emerging from learning.
- This is old work, and similar nets were used to categorise
yeast (52%), and other tasks.
- It's pretty general.
- What we'd like is the correct category neurons firing
for an input.
