To the practitioner, the most important skill is not in machine learning. The most important skill is in understanding the problem. Implicit in this statement is that the practitioner should also at least understand which machine learning algorithms would be suitable for the problem at hand. Obviously this entails understanding how the machine learning algorithm works.

New people in the field often ask me whether deep learning will solve all their problems. The answer is emphatically no. The solution must be tailored to the problem. Indeed, often, non-deep-learning solutions outperform deep learning solutions in terms of speed and accuracy. These are typically simple problems, so that's a good rule of thumb there: if the problem is non-compositional, you most likely do not need to use deep learning.

What do I mean by non-compositional? Recall from Chapter 1, How to Solve All Machine Learning Problems, when I introduced the types of problems, and how problems may be broken down into subproblems. If the subproblems are themselves composed of further subproblems, well, that means the problem is composed of subproblems. Problems that aren't compositional do not need deep learning.

Granted, this is a very gross overview of the issue. A finer understanding of the problem is always required.