The digits here are written in white pixels over a black background. I don't think anyone would have a problem with identifying a digit if it was written in black pixels over a white background instead. As for a computer algorithm, things are a little different. Let's train our classifier as usual and see whether it will have any issues if the colors are inverted. We will start by training the algorithm on the original images:
clf = KNeighborsClassifier(n_neighbors=3, metric='euclidean')
clf.fit(x_train, y_train)
y_train_pred = clf.predict(x_train)
We then create an inverted version of the data we have just used for training:
x_train_inv = x_train.max() - x_train
The nearest neighbors implementation has a method called kneighbors. When given a sample, it returns a list of the K-nearest samples to it from the training set, as well as their distances from the given sample. We are going to give this method...
