Experiment 10 is from our Hierarchical CAs paper pg.7. The network weights are derived from a Hopfield net that we calculated. Uses ParametersExp10.dat
When the net is loaded, it has 4 stable states. Note that the net consists of 40 identical sets of 10 neurons. The weights have been calculated to insure that the 4 stable states (dog, cat, rat, and mammal) exist. These correspond to neurons 0-3-4-5-8 dog, 1-3-4-6-9 cat, 2-3-4-7-8 rat, and 3-4-9 mammal.
If 200 neurons are activated, all of the basic category neurons stay active for all 50 cycles. If the input pattern is switched to Mammal (pattern 12) and 120 neurons are presented, they all remain on.
The default is 40 neurons externally stimulated, and this leads to a fair number of neurons activated, but not all 200. This can be improved by generating a new network. This network has an extra connection from each neuron to its corresponding neuron in the next version duplication. (E.g. 0 is connected to 10, 11-21). This can be done by changing the generate network type item on the parameters menu, and generating a new network on the net menu.
This new network has moved away from a hopfield net by introducing uni-directional connections. It also improves the behaviour by activating more neurons from a smaller input. Pearsons measurements can show this by setting the set correlation step values and using the print correlation menu item all on the measurement menu.