Before we can talk about concrete machine learning algorithms, we have to talk about how best to store the data we will chew through. This is important as the most advanced learning algorithm will not be of any help to us if it will never finish. This may be simply because the mere process of accessing the data is too slow, or maybe its representation forces the operating system to swap all day. Add to this the fact that Python is an interpreted language (a highly optimized one, though) that is slow for many numerically heavy algorithms compared to C or Fortran. So we might ask why on earth so many scientists and companies are betting their fortune on Python even in highly computation-intensive areas.

The answer is that, in Python, it is very easy to offload number crunching tasks to the lower layer in the form of C or Fortran extensions, and that is exactly what NumPy and SciPy do (see https://scipy.org). NumPy provides the support of highly optimized multidimensional arrays, which are the basic data structure of most state-of-the-art algorithms. SciPy uses those arrays to provide a set of fast numerical recipes. Matplotlib (http://matplotlib.org) is probably the most convenient and feature-rich library to plot high-quality graphs using Python. Finally, TensorFlow is one of the leading neural network packages for Python (we will explain what this package is about in a subsequent chapter).