Alternatives to Random Permutation

[georgwiese]

How else can one preprocess data: other approaches to thermometer encoding / random permutation? They could improve accuracy.

• Read literature on WNN techniques which were considered [1 day]
• Outline the techniques & effects [1 hour]
• Think of some alteration / other approaches [1 day]
• Try the alternative data preprocessing - report results [4 hours]

[georgwiese]

I think it might make sense to separate how pixels are encoded (-> thermometer encoding) from the wiring of the network (-> random permutation).

A few thoughts:

• Pixel encoding:
  • I think "feature engineering" is the right key word here. It was done heavily before DNNs became the go-to way to do image processing (they consume the (normalized) pixel values directly).
  • For example, the Sobel filter could be used to preprocess inputs and detect edges.
  • Perhaps there are also more global features, i.e., not on the pixel level?
  • Note that any feature engineering would have to be proven also by the circuit.
• Wiring:
  • I guess there could be some kind of windowing, similar to CNNs?
  • But then I'd be worried about the WNN looking for exact matches. Perhaps it should be replaced with something like: a discriminator is active if the maximal hamming distance to the closest training example of that class is low (but not necessarily 0).