As we discussed, the more traditional predictive models such as linear regression don't scale well, because they always need to calculate the whole solution using all the available points or data. These types of techniques or models have no ability to remember, learn, and improve, and they are generally classified as supervised models. This has led to the evolution of more advanced learning models known as reinforcement learning (RL) techniques for solving ML problems. In fact, deep learning and deep reinforcement learning techniques now outclass statistical methods in performance and accuracy by several orders of magnitude. However, that wasn't always the case, and statistical methods are also improving just as dramatically everyday. It really is an exciting time to be getting into Machine Learning.

The following diagram demonstrates the reinforcement learning process:

Reinforcement learning process

In the diagram, you can see that there is an Agent (assume computer) and the Environment (game or real world). The Agent acts on Observations from the Environment, and those actions may or may not be based on Rewards. An RL system using rewards is known as reinforcement learning. The learning method we will use in this chapter is called supervised learning since we are labeling or training to a specific output class. Unsupervised learning is a class of training that doesn't label data but just uses techniques to classify or group data.

Part of the problem in defining all these ML concepts is that they often get woven together, where one learning algorithm or technique is layered on top of another, perhaps using RL with or without supervision. It is quite common, as we will see, to use multiple different layers of techniques to produce an accurate answer. This layering also has the benefit of being able to try multiple different approaches quickly or swap a technique out for something better later.

Deep learning is the term we use to describe this layering process. DL can be trained using any of the training methods we talked about. In any case, we need to stop talking in generalities and actually look at the DL process.