CNN on MNIST

Best results:

Effect of dropout:

Dropout though important is not most critical in changing accuracy.

Effect of Kernel size:

Kernel size of 3 had best accuracy. 5 was too big for 28x28 image and 1 too insignificant.

Stride effect:

Stride had significant effect.
Stride of size 1 ensures that information is not deleted from image. For that part there were max pooling layers.