I want to build a KNN classifier while constraining the number of training samples. For example, I have the MNIST dataset with a total number of 60K training samples and I\'