Myths of machine learning

The argument goes, if something has never happened before it’s predicted probability must be zero, what else could it be? That’s false as events are composed of many smaller components, all of which have relationships and similarities. The power of machine learning is identifying these relationships, and using them to predict rare events with high accuracy.

Experts in different fields have invested significant human effort to develop domain knowledge. It’s easy to think that this knowledge is lost when a machine learning system is used. A vital part of machine learning is feature engineering, the process of extracting patterns or features from raw data. Domain experts are often better than machines at suggesting features that hold predictive power. As such, domain experts form a key part in defining the input to a machine learning system, from which preexisting knowledge can be extracted, extended, and refined.

In supervised learning an algorithm is given an input and a desired output, and the aim is to learn the mapping from input to output. Think of it as a school test, where there are questions and answers, and you are graded by how close your answers are to the actual ones.

Now imagine there are no answers, what can you learn if you only have the questions? In unsupervised learning the aim is to uncover the underlying structure or distribution of the data in order to learn more about it. For example, finding groups of users that behave in the same way, or identifying events that are anomalous. Unsupervised learning can even generate completely new data. If given enough images of faces, we can train a computer to create realistic images of people who don’t even exist.