Zero centering bug in "Training and Optimizing Neural Networks"
I think there's a bug in Training and Optimizing Neural Networks.ipynb
These lines of code in section "4. Weight Initialization" undo the zero centering. All subsequent networks are then trained with uncentered training data.
X_train_zc = X_train.astype('float32')
X_test_zc = X_test.astype('float32')
I think it should be
X_train_zc = X_train_zc.astype('float32')
X_test_zc = X_test_zc.astype('float32')
This problem occurs in other notebooks as well.