2013-05-14 #1

Perceptra 2D Test Run – 1 14th May 2013

These 2D versions do not include borders. The first one uses the update name classes that makes a binon ambiguous when both the target classes are different.  It does not assign the same name class to both target class pointers like the second version. Thus it flags more lower level classes of rectangle combinations as ambiguous and creates more higher level combinations than the second version. Higher level combinations are more likely to match / predict correctly.

The second version gets an even higher no prediction rate. The no predictions are caused by the combination of rectangles that are found in the figure all being ambiguous.

There are also lower prediction rates for digits that don't have holes in them. For example the digit 1 if considered as a tall rectangle (level-2) can be found as part of many other digits. It therefore gets marked as ambiguous because it is found inside other more complex objects. I think a solution might be to only update the target class links to the name binon for those classes / binons that are not reused as part of a higher level binon.