Convolutional neural network for hand-written digit recognition
Input
28x28 pixels
Convolution 1
8 filters each 5x5 pixel
activation: ReLU
output: 8 images each 24x24 pixel
MaxPool 2
2x2 pixel segmentation
output: 8 images each 12x12 pixel
Convolution 3
16 filters each 5x5 pixel
activation: ReLU
output: 16 images each 8x8 pixel
MaxPool 4
2x2 pixel segmentation
output: 16 images each 4x4 pixel
Dense 5
Full Connection (2560 weights)
activation: softmax
output: 10 neurons (values between 0 and 1)
Loss function: Categorial Cross Entropy
Optimizer: Adam