Experiment 2 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. As this is repeated, axonal strengths are increased, and eventually, the network persists for all of the additional 30 cycles. Uses ParametersExp2.dat
At the first cycle, a random network has been generated. There are inhibitory and excitatory neurons, but all of the axonal strengths are near 0. Initially 100 (of the 400) neurons are externally stimulated. These will be stimulated for 20 cycles. Note that after a few cycles less than 100 (around 70) are firing; this is because some of the neurons have fatigued, and the external stimulation is insufficient to fire them.
After 20 cycles, external stimulus ceases for 30 cycles. On the 50th cycle, a new pattern of 100 neurons is selected, and the process is repeated.
A correlatory Hebbian learning rule is used. So when two neurons are coactive, the synapse between them is strengthened. (If the pre-synaptic neuron is active, and the post-synaptic neuron is inactive, the synapse is weakened.) As the initial weights are very low, the strength has increased after 50 cycles. At 71 cycles, neurons are still firing without external activation. After 140 or 190 cycles neurons are still firing because the strength has increased so much.
This shows that a simple CA can be learned and can persist without external stimulus.