2013-05-13

Perceptra 2D Test Run – 1 13th May 2013

This is without border rectangles. Note the large number of no predictions. It did quite well on some numbers but not others.  It had a serious problem recognizing ones. This may be because they are very simple and the vertical rectangle that they contain can be found as part of other digits. Thus these vertical rectangles become ambiguous and contribute to the no prediction total.

Digits with holes in them such as the zero, four, six and eight are more successful.

The images didn't too well compared with the algorithm that uses the borders. The no predictions are higher. And the digit seven was the most difficult to recognize.

For 16x16 bitmaps the ones and fives were the hardest to recognize.

What I could do is recognize those cases where all the rectangles were used to form one highest level class and then change its name binon to be unique (not ambiguous). Could updating the name binon links such that we always lose the oldest one do this?