CANT 2.0 Experiment 11: Hierarchical Cell Assemblies
Experiment 11 is from our
Hierarchical CAs paper pg.12. Hierarchical CAs are formed.
.
This is a 20x20 nets, and is presented three types of patterns.
We've called the patterns dogs, cats and rats. These have
overlapping features. After training
each of these are activated, but the presentation of a
higher order pattern, Mammal, is also activates that CA.
Uses ParametersExp11.dat
Training
When the system is loaded up, it is training mode. In this case
training is done without spread of activation and each pattern
is presented for one cycle. While the basic patterns consist of
200 neurons (and the higher category consists of 120) only 40 neurons
are presented, both in training and testing. Presentation for
one cycle without spread of activation is not biologically viable,
but reduces the likelihood of poor stable states forming. Training
should run for roughly 2000 cycles, though other times will produce
reasonable results.
Testing
Use the step button to stop the system. Turn learning off, and
turn Change pattern each time off (both on the Run Menu). Now
a pattern will be presented for 20 cycles and will be allowed
to continue for 50 cycles. Run this for a say 350 cycles.
You can use Print Correlation to display the Pearson's values.
For example two Dog patterns (5 and 155) are highly correlated.
(You can set these by setting the set correlation step 1 and 2
fields. These are all in the measure menu.) Dog and Cat (e.g.
5 and 55) have a quite low correlation and maybe even a negative
correlation. Dog and Rat are more closely correlated, but still
differ from an intra-category comparison.
Testing the Higher Category
Change the input pattern so that it is presenting a Mammal. Do this
by the Select Pattern item on the run menu. Set the pattern to 12.
After running this a few times you can compare versions of Mammal to
each other and to the basic category using Print Correlation. Again
you will note that it Mammals are correlated highly with each other
and correlated with the basic categories but less so.
Page last modified