sindhri / MNIST

Use 28x28 pixel images in the MNIST dataset to classify digit 0-9 via deep learning: MLP vs. CNN

MNIST digit recognition

Use 28x28 pixel images to recognize digit 0-9
Credit: Deep Learning with Python by Jason Brownlee

baseline MLP

flattern and normalize the input
MLP with two dense layers
categorical cross entropy as the loss function, adam as the optimizer, and accuracy as the metrics.
10 epochs with batch size = 200
Baseline Error: 1.67%

simple CNN

reshape and normalize the input
simple CNN with 1 convolutional layer, 1 pooling, 1 dropout, 1 flattern, 2 dense layers
Much slower than MLP
categorical cross entropy as the loss function, adam as the optimizer, and accuracy as the metrics.
10 epochs with batch size = 200
CNN Error: 1.09% reduced error rate

larger CNN

larger CNN with 2 convolutional+pooling layers, 1 dropout, 1 flattern, 3 dense layers Much slower than MLP
categorical cross entropy as the loss function, adam as the optimizer, and accuracy as the metrics.
10 epochs with batch size = 200
Large CNN Error: 0.77%, improved!