We know that computers are good at yes and no answers, or 1s and 0s, if you like. This means that a computer fundamentally cannot reply with an -ish answer, so it cannot say yes-ish to a question. Bear with me for a moment, as this will become clear shortly.

At its most basic level, we can say that machine learning boils down to teaching computers to learn in the same way that we do. They learn to interpret data from all sorts of sources and use this learning to classify that data. The machine will learn from successes and failures, which will, in turn, make it more accurate and capable of making even more complex inferences.

Getting back to the idea of computers working with yes or no answers, when we come up with an answer that amounts to well, it depends, we are largely coming up with multiple answers based on the same input—the equivalent of multiple routes through to yes or no answers. Machine learning systems are becoming much better at learning, so the algorithms behind them are able to draw on more and more data, along with more and more reinforcement to make deeper connections.

Behind the scenes, machine learning applies an incredible array of algorithms and statistical models so that systems can perform set tasks without having to be given detailed instructions on how to accomplish those tasks. This level of inference is light years away from the way we have traditionally built applications, and this draws on the fact that, given the right mathematical models, computers are very, very good at spotting patterns. Along with that, they are doing a huge number of related tasks simultaneously, meaning that the mathematical models underpinning the learning can take the results of their calculations back in as feeds to themselves in order to build a better understanding of the world.

At this point, we must mention that AI and machine learning are not the same. Machine learning is an application of AI based on the ability to automatically learn without being programmed to deal with a particular task. The success of machine learning is based on having a sufficient amount of data for the system to learn for itself. There are a number of algorithm types that can be applied. Some are known as unsupervised learning algorithms, while others are known as supervised learning algorithms.

Unsupervised algorithms take in data that has not been classified or labeled previously. The algorithms are run on such datasets to look for underlying or hidden patterns, which can be used to create inferences.

A supervised learning algorithm takes its previous learning and applies it to new data using labeled examples. These labeled examples help it learn the correct answers. Behind the scenes, there is a training dataset that learning algorithms use to refine their knowledge and learn from. The greater the level of training data, the more likely the algorithm is to be able to produce correct answers.

There are other types of algorithms, including reinforcement learning algorithms and semi-supervised learning algorithms, but these are outside the scope of this book.