Ahead of our event ‘Learning machines – how computers got smart’ on 12 April we caught up with Professor Marcus du Sautoy OBE, event chair and member of the group of experts leading the Royal Society Machine Learning policy project, to get his thoughts on machine learning, present and future.
Can you tell us a little bit about what machine learning is?
Machine learning is an approach to computer programming, which creates algorithms that learn from data and self-improve. In the past, programmers would tell a computer exactly what to do to solve a problem; with machine learning, it is now possible to write a computer program that adapts when it encounters new data. In effect, it learns to solve a problem. Machine learning requires a human to create the algorithm, but the program enables the machine to learn from its interactions with the environment.
For example, in 1997, IBM’s computer DeepBlue defeated the world chess champion. This machine was programmed to analyse as many possible chess moves ahead of play, score them and then take the appropriate move. This approach did not change over time. In contrast, Google DeepMind’s AlphaGo recently beat the world’s top player at Go, a game with far more possible moves than chess. Unlike DeepBlue, AlphaGo played millions of games against itself to learn from its mistakes and change its approach accordingly. This is an example of machine learning in action.
Machine learning is possible now because of the availability of large data sets. The age of ‘Big Data’ means we can let computers improve in an evolutionary way as a result of training on large datasets, whereas before we programmed computers without access to the data they might encounter.
How is machine learning used at the moment? What applications might people already be familiar with?
Most people will already have encountered machine learning in some form; for example, it is used in search engines and in recommender systems, which give suggested preferences for services and products. We might not think about the technology behind these applications, but it is hard to write an algorithm that can anticipate preferences! If an algorithm learns what people might like based on previous choices, it can make very accurate recommendations. Examples of this can be found in Netflix, Amazon and online dating services.
Machine learning is also used in navigation systems, to help us get from A to B. The more data we give to machines by the journeys we make, the more they are able to adapt and learn more routes. This dataset develops as we make different journeys and choices, so the algorithm has to update and change itself.
What application of machine learning do you use most?
I suppose through my interaction with the internet and by agreeing to cookies which store my data, the computer is learning what I am interested in. In a sense, the computer is building up a personal relationship with me. I also use navigation systems regularly.
What do you think might be the key applications for machine learning in the future?
In any area where there is a lot of data, machine learning will be able to be put to use.
For example, while it is fun to watch IBM’s Watson playing Jeopardy or AlphaGo playing Go, both of these companies are now looking at uses beyond gaming. Healthcare might be a field where these machines can be put to use, by learning from interactions with patients and making recommendations for treatment. Machine learning could also be put to broader use in navigation, through applications in driverless cars, and could help develop public policy or improve public services.
I am excited to see what role machine learning will have in science. If we look at CERN, for example, this is a project where a vast amount of data has been generated. However, we are currently only using this data to look for things at CERN that we are already expecting to find there due to theoretical considerations, such as the Higgs-Boson particle. But what if there are patterns in this data which we haven’t seen and which we might not have anticipated? Machine learning could highlight these and add insights to this research.
In my field, mathematics, we are excited about machine learning. AlphaGo’s recent success showed that a computer can demonstrate what we might think of as imagination or creativity, as well as following logical sequences. This is similar to what we need to do to find proofs in mathematics. So machine learning could help mathematicians to find proofs in the future.
Why do you think it is important for the Royal Society to be doing a project about machine learning?
The Royal Society is currently looking at machine learning, and the benefits, risks, challenges and opportunities associated with this technology. As this technology has such a broad array of potential applications, some of which will have higher stakes than others, it could have a significant impact on the economy and society. So, the Society is interested in looking at the different uses of machine learning, and how this technology can be developed in a way which is beneficial for society.
AlphaGo demonstrated that machines could make moves that looked like mistakes to human players, but were based on seeking out patterns unseen to us. This highlights some interesting questions about how far we would be willing to trust predictions from machines, especially if it is hard to unpack why a machine has chosen a particular course of action, based on a learning process we haven’t been involved with. These issues need to be considered as the technology develops, and the Royal Society’s project will be driving such discussions.
What are you looking forward to most at the 12 April event?
I am looking forward to hearing the stories of the three speakers. For example, I visited Chris Bishop’s laboratory recently to see how machine learning is being used. One research strand is developing a camera that can identify a moving body, having interacted with a range of images of people with varying shapes and sizes. This is essentially a form of ‘computer vision’ as, amazingly, the machine can recognise what is a body. There are exciting applications of machine learning in the field of robotics, which Maja Pantic and Sabine Hauert will be able to tell us about. For example, I’ve seen previously how two robots can learn a new language by interacting with each other.
I am also looking forward to the interactive exhibition, where everyone will be able to experience machine learning in action and quiz the developers.
I also really want to hear from the public about their feelings on machine learning. How much do people know about the use of machine learning in our day to day lives already? Are they excited or nervous about the future of this technology?
We’ve been through an AI winter, where the expectations which surrounded this technology in the 1970s came to nothing, but now, with the increased availability of data, we are in a golden age for machine learning; I think this technology will be a very exciting part of our future.
‘Learning Machines – how computers got smart’ will take place on 12 April. Keep your eyes on In Verba for further information about the Royal Society’s machine learning project.