Compensatory Learning
- Correlatory Learning is sufficient to learn orthogonal patterns.
- Orthogonal patterns are the kind we've used where now patterns
share any neurons.
- However, we've found that overlapping patterns are much easier
to learn if a compensatory learning rule is used.
- Compensatory means that the total synaptic strength from a neuron
is forced toward a value.
- To make this work, switch on the Execution/Compensatory Learning
menu item
- The goal strength is set by Set Parameters/Saturation Base.
- It reduces the likelihood that neurons that are in multiple
patterns will cause the patterns to be treated as one.
- It also an make learning faster.
- Change the param.xml file so that two of the patterns overlap.
- Using the correlatory mechanism, it's easy to get 3 patterns
instead of 4.
- Using the compensatory, it's easier to separate them.
- Try this both ways.
