mnist-from-scratch MNIST done with paper, pen, pencil and numpy. single hidden layer with 32 units no preprocessing, no regularisation normal random initialisation train/val accuracy: 93% Backprop derivation :)
