Experiment 3 is from our 1999 Tech Report pg. 12. A 20x20 net is presented with varieties of a pattern for 20 cycles, then allowed to run for 30 more. It is then presented with a different sort of pattern. These types are 100 of the top 200 neurons, or 100 of the bottom 200 neurons. Uses ParametersExp2.dat
After a few iterations of the top and bottom pattern, the neurons in the top will continue to fire after external stimulus in the top has stopped, as will the bottom neurons. When a top pattern is presented, neurons in the bottom will never fire, and vice-versa.
This is due to the Hebbian learning rule. Initially, all of the weights are random and small. So a synapse between a neuron in the top and one in the bottom has the same weight (roughly) as that between two neurons in the top. However, the synapse between the two neurons in the top will strengthen, because they are coactive, while the one between the patterns will weaken becuase they are never coactive.
This leads to two independent CAs. Virtually any pattern will form CAs if they have enough neurons. It could as easily have been the left 200 and right 200 neurons.