A Tour of AI

Artificial Intelligence has been making headlines recently. Over the past year an AI has beat the worlds best Go and Poker players. Regular talk of driverless cars fill social media and news websites. These are all things that people have previously said would be impossible for a machine to do. We are in the middle of a boom in interesting AI applications. Looking in from the outside it's easy to feel like AI is just magic. I decided to put this article together to help myself and others feel their way through the fog of buzz-words. Here is what I will be covering.

Strong & Weak AI

What exactly does Artificial Intelligence (AI) mean? An AI is any machine which has capabilities you would normally associate with an intelligent being. AI is often divided into two categories; strong and weak. A strong (or true) AI is a machine that is at least as capable as a human being in all ways. Think terminator, Chappie, HAL, Wall-E, etc. Obviously, there aren't any strong AIs yet. Weak AI refers to the kinds of AI we have today that can perform some specialised function. Technically all modern computer hardware and software is AI, but we only really refer to something as an AI if it surpasses our expectations of machines. As new kinds of AI become common they start to be referred to as AI less and less. You probably wouldn't refer to a chess game as AI for example, but in 1997 when IBM's chess AI Deep Blue beat world chess champion Gary Kasparov it was considered the height of artificial intelligence.

Machine Learning & Symbolic AI

The term machine learning has been thrown around a lot recently. In 1997 Tom Mitchel defined machine learning in very practical terms:

"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E."

Which is just a fancy way of saying that any machine that gets better at some task with experience uses machine learning. Machine learning has largely been responsible for the boom in AI applications we are seeing today. More often than not machine learning simply uses statistics to work out patterns in data, and then uses these patterns to make assumptions about new data that the algorithm hasn't seen yet. Machine learning has been around for a while, but it wasn't until the 80s that people started to consider it a real and viable method for creating AIs.

Before that the prevalent method for creating AIs was to use symbolic AI. Symbolic AI is a term you may not be familiar with, but essentially it is the creation of AI through defining rules and patterns that a machine must follow. In symbolic AI the engineer defines the rules of how the machine works. With machine learning the engineer tells the machine to find the rules itself.

Most of the hype around AI recently has been from machine learning, because it has become clear that symbolic AI is impractical for pushing past what we've already achieved. Defining long lists of rules to follow is a long and painful process fraught with errors. It may well be impossible to write rules for every possible scenario for an intelligent system. The rest of this article focuses on terms from machine learning.

Supervised & Unsupervised Machine Learning

People very often break machine learning into two broad categories; supervised and unsupervised learning. At the heart it, the different between these two categories is quite simple. In supervised learning you provide the machine with the right answers for the examples it uses to learn. For instance, you may give an AI pictures of cats and dogs and tell it which one is which. You could then present it with a new picture and ask it whether it is of a cat or a dog. This would be supervised learning. In unsupervised learning you provide the machine with the pictures of cats and dogs but don't tell it which ones are which. In this case you might ask it to split the images into two categories and it might split them into pictures of cats and dogs, or it might choose some other category like pictures taken indoors and outdoors.

In reality AIs often don't fall neatly into either supervised or unsupervised learning because they may use algorithms from both.

Clustering & Anomaly Detection

Clustering and anomaly detection are unsupervised learning techniques. Clustering is the process of taking a set of things and grouping the most similar ones together. Our unsupervised learning example of grouping similar photos together is exactly this, except it wouldn't have to be into two groups. It could be any number of groups. Anomaly detection is used to find unlikely (or anomalous) data in some set of entries. Typical examples include fraud & hacking detection, fault detection in manufacturing, and detecting weird data for academic purposes.

Classification & Regression

Classification and regression are terms generally used to described types of supervised learning. Classification is when you try to determine whether something fits into one class of thing, or another. For example the cat and dog example we used previously would be a classification algorithm. It takes pictures and classifies them as a picture of a cat, or a dog. Regression problems on the other hand try to produce fluid values. Imagine a program that you give the details of your house to. It takes the number of rooms, floors, bathrooms, en-suites, as well as the size of the rooms, and the style of the house, etc. It uses those details to produce a price estimate for your house. This would be a regression algorithm. Other examples include predicting the temperature, sales of a product, or the value of stocks.

Specific Learning Algorithms

Supervised, unsupervised, classification, regression, clustering, and anomaly detection algorithms are all categories of algorithms. There is an extensive list of specific learning algorithms each coming with their own name, pros, cons, and specialist uses. One of the most highly talked about is Artificial Neural Networks, which are classification algorithms. Although the workings, pros, and cons of each algorithm are beyond the scope of this article below is a list of common learning algorithms and which categories they fall into.

The above are limited to algorithms I have some basic understanding of, but there are an enormous number and variety of machine learning algorithms. Some of the above can also be split into various sub-types, again with their own use cases, pros, and cons.

I hope you've found this illuminating. At some point in the near future I would like to write up an example implementation of one of the above. In the mean time, if you have any questions or criticism post it in the comments below.