Experiment 5 is from 2001 Tech Report pg. undone. This is a 20x20 nets, and is presented two types of patterns, neurons from the top half or neurons from the bottom half, but the middle two rows appear in both. Uses ParametersExp5.dat
In this experiment, the presented patterns overlap by 10% of the neurons. The problem comes from these overlapping neurons that have connections (justifiably) to both CAs and are thus members of both. It's simple to prevent the non-overlapping neurons (parts and 1) from activating the opposite pattern, because during training they are never coactive and thus the excitatory synapses go to 0. However, the overlapping neurons (part 2) should have smaller weights to the non-overlapping neurons than to each other.
Compensatory learning allow the weights to grow beyond correlation. Correlatory and compensatory learning both keep the 2-1 weights half the 2-2 weights and half the 1-1 weights.
One way the top is kept separate from the bottom is via inhibitory links. Both are ignited, but both inhibit the other, so only one comes on at a time.
Note that sometimes just one CA forms. In this case you can simply make a newnet and train from scratch.
Load the system up, and train. The ones I've been running have taken around 2000 cycles, but longer tends to activate more neurons during ignition. Note, the bottom variant comes on at cycle 0 and 50 etc. while the top variant comes at 25, 75 etc.
Use the print correlation menu item to see how similar the CAs are. Set Correlation 1 to 10, and correlation 2 to 60, and you should get a positive correlation (say .40) while 10 and 35 get a negative correlation (say -.05). Note that these values are mod 1000, so the above comparison works for 2010 and 2060. If you want to compare 1910 to 2010 and you're at step 2100, set correlation 1 to 910, and correlation 2 to 10.