Machine Learning: A Well-Dressed Emperor

Source: Deep Learning on Medium

In What is Machine Learning? I have given an informal definition of Machine Learning:

Machine Learning happens when a machine, over a period of time, gets better at a task it wasn’t explicitly told how to do.

Now, such a task can definitely be labeling stuff. One could give a computer images of cats and images without cats, and tell the computer which ones contain cats and which ones don’t. Using Machine Learning techniques, the computer could then learn from these examples and correctly give a new, unseen image the label “Cat” or “No cat”. That’s Supervised Learning, and it appears that Cassie equates Machine Learning with this in her article.

Machine Learning, however, is broader than “just” Supervised Learning. Reinforcement Learning, for example, is an area of Machine Learning where computers can learn from “examples” not given by humans. For example, a computer that’s given the rules of Connect Four could play a million games of Connect Four against itself. It could then learn to play better.

The computer learns from examples it generated itself.

We have to tell the computer that winning situations are good, and situations before a winning one a little less good (since they led to a winning situation), etc., but that’s only some basic rules: we don’t give the computer actual examples of board states we think are good. We tell it that four in a row means winning (or losing), and the computer applies this logic to board states it generates itself (by playing against itself). The same goes for other board states it generates: the closer they are to a winning board state in that game, the higher the score it will attach to that board state. That gives extra examples to learn from. That way, the computer learns from examples it generated itself.