Nearly everyone who keeps up to date with technological developments has heard of the term ‘machine learning’. It’s a nice buzzword that is being used more and more in recent years. But what does machine learning actually entail? What is the technology that lies behind it?
To coincide with the launch of Clocktimizer’s machine learning engine, we will be releasing three blogs written by Susan Brommer. Susan is Clocktimizer’s resident AI genius and the person responsible for building much of our own machine learning engine. In the blog series we will explain what machine learning is, how it came to be, and what we can expect from it in the future. We will even go into some detail of the inner mechanisms of machine learning. Last week we discussed why machine learning is important. In this second blog we are going to take a look at how machine learning evolved over the years, and what the big achievements are.
In the beginning….
… there was Alan Turing, an English mathematician, born in 1912. You may know him from his essential role in ending the Second World War, by developing methods to break the German ciphers*. We, however, start the history of machine learning with Alan Turing because he came up with an experiment that can decide whether a machine is intelligent.
This is the so called Turing test, and involves two humans and one computer. One of the humans is the interrogator, sits alone in a room, and has the possibility to communicate via two computers. One computer lets the interrogator communicate with the other human. The other computer lets the interrogator communicate with the machine. The interrogator does not know which computer communicates with whom. Turing stated that we can call a machine intelligent, if the interrogator is unable to differentiate between the human and the machine based on the conversations he has via the computers.
Of course, there are several questions and critiques regarding Turing’s test, but it did give rise to the desire to create intelligent machines. Researchers then used knowledge based algorithms: algorithms that are fed simple rules from the real world and act upon this ‘knowledge’. This led to some success, like a tic-tac-toe winning computer in 1952. One could argue, though, that while this approach may result in intelligent machines, it does not really involve learning. The first machine that actually was self-learning was created in the late ’50s, by Arthur Lee Samuel at IBM. He created a checkers playing program that improved on every game it played. Not unsurprisingly, Samuel is the man who coined the term ‘machine learning’ in 1959.
Winter is coming
Researchers were optimistic about the developments within machine learning. Unfortunately, in the early ’70s, enthusiasm for artificial intelligence decreased dramatically, as it failed to meet expectations: intelligent algorithms proved useful only in very specific contexts. Most of the time they were unable to handle unusual (real-world) inputs, and it was very expensive. The research field went on to suffer several financial setbacks. Accordingly, research on artificial intelligence, and thus on machine learning, slowed down. This downfall is known as the AI winter, and covers about twenty years.
In the early ’90s, new life was breathed into machine learning research. Whereas before the focus had been knowledge based algorithms, now research engaged in data based algorithms. Opposite to knowledge based algorithms, where computers were fed rules about the world, data based algorithms are fed data about the world, and extract the rules about the world themselves from this data. This new focus on data was made possible by development and refinement of statistical methods, a huge increase in digitised information, and fast distribution of data via the internet.
Games, games, games
With the new focus on data, came a new focus on solving problems in a practical matter, in particular games. Some of the milestones machine learning has achieved through playing games include: in 1992 machines were able to rival human backgammon players; in 1997 IBM’s Deep Blue beats Kasparov, the world champion at chess; in 2010 IBM’s Watson beats two human champions in Jeopardy!; and in 2016 Google’s AlphaGo becomes the first computer to beat an unhandicapped professional human player.
Aside from machines that learn to play games, machine learning became a bit of a game itself. In 2006 Netflix launched a competition aiming to improve its own recommendation algorithm by ten percent. The prize was won in 2009. In 2010 Google founded Kaggle, a platform for machine learning competitions. Those of you wishing to help train a machine learning engine can even head to Quick Draw and see if a neural network can recognise your doodles.
Games are an easy target for machine learning research, since the rules often are very clear, and it is easy to measure how well an algorithm performs. However, machine learning has also managed to learn more human tasks. Nowadays, machine learning beats humans at identifying individuals on pictures, interpreting human speech, detecting cancer in human tissue images, translating dozens of languages, and lip read from television videos. It also can recognise cats in YouTube videos.
Are they superhuman?
If you list all these achievements, you could almost say that machines are better at being human than humans themselves. But this is not entirely true. They might be better at certain tasks previously only performed by humans; however they are still limited to doing that specific task only. You could argue that machines will therefore never make a real impact on the world by themselves. For that, it is essential to work together with the machine. An algorithm can search through thousands of gigabytes of data for you, and then you are the one to interpret and act upon this result. Where does this result come from? What does this result mean? How can we use this result? Machine learning algorithms are there to give you new knowledge. Applying this knowledge to make a change is still a task reserved for humans.
Something you could also ask is whether the machine really understands what it is doing? Does a chess playing machine really understand that it is playing a game? Does it really want to win? Do computers actually learn? Do they gain knowledge? Or does a machine just get pieces of information, and respond in a way that its algorithm tells it to, without understanding what it is doing?
These questions are a bit philosophical, but it is important to realise that computers are not per se superhumans. Most of us feel that there is still something that distinguishes machines from humans, however difficult it is to pinpoint what exactly that is. When using machine learning algorithms, you should bear this in mind. The machine might help you get information, but it is still up to you to make decisions on what you do with this information.
What does the future hold?
What can we expect from machine learning in the near future? Obviously, machines will become more powerful, will be able to process more data, and will give more accurate results. Also, we should expect computers to become better at communicating with humans. A few years ago, we had to learn how to interact with a computer; nowadays, computers are learning how to interact with humans.
More and more companies will feel the need to use machine learning. Many companies acknowledge the power of machine learning, have a lot of data, and want to get value from it. Whether it be to improve customer experience, to predict pricing, or to optimise workflow: if there is data, machines can learn from it.
As an answer to the increase in demand for machine learning, more businesses will start to offer machine learning as a service (MLaaS). This allows more businesses to easily use machine learning, without them having to implement such algorithms themselves.
Do not buy a car without knowing a bit about how a car works. (Sound familiar? This was the motto of our Coding Class at Legal Geek!) The same holds for machine learning: if you go hunting for a machine that can learn from your data, you should at least know a bit about the inner workings of a machine learning algorithm. In the third and last part of this blog series, you will learn the basics of machine learning.