So, I tried to test out exact match convolution, which can be defined as:
It consists from three parts:
- pixel offset
- standard convolution
- convolution bias
If you test it out with DeeperThought, with config:
where convolutionsq is exact match convolution (emc), you get pretty good results:
Top accuracy on test data: 98.68 % (Dataset - MNIST digits).
Not as good as with convolution (up to 99.41 % accuracy), but pretty close. Are we really just looking for exact matches and we do so by convolution? Hint: Linear Regression on a Set of Selected Templates from a Pool of Randomly Generated Templates.