MultiMNIST with CNN

Best results:

"Complete match" accuracy:

It reaches total 80% at max, which is huge as compared to MLP for same task.

Other comparisions:

Stride 1 gives almost all good accuracies. I makes sense as our model also has pooling in it.
Higher kernel sizes give better results. This is reasonable as we have 64x64 image now, and only 2 pooling is not enough to summarise features all over image. Larger kernel also result in more complex patterns detected.
Smaller lr is better.
Dropout is prefered but <= 0.3 is optimal.