Experiments with bisector models

[dgoldman-pdx]

Pursuing an idea, which so far hasn't panned out...

Because pure-MLP models execute ~35 times quicker than convolutional models, attempt to create a set of smallish pure-MLP models that together add up to a reliable digit recognizer.

Attempts so far have involved creating 4 different models, each of which bisects the set of digits {0...9} into a unique pair of subsets (each containing 5 possible digits). These bisectors have been defined so that if all 4 models perform correctly, then only a single digit will belong to all 4 chosen subsets.

So far, models "B" and "C" seem to work pretty well, but models "A" and "D" make a lot of mistakes.

Further experiments might involve:

• making better bisector models;
• using more models (say, 6 at least), to compensate for inaccuracies in any one or two;
• using models that divide {0...9} into unequal-sized subsets, or more than 2 subsets.