Professor Zoubin Ghahramani explores the fascinating world of machine learning and leads us to the invisible algorithms underlying many of the tools we now use every day.
Professor Ghahramani is speaking at this year's Festival during the event Intelligence and learning in brains and machines.
CSF: What exactly does machine learning mean and how does it work?
ZG: Machine learning is the field of research dedicated to understanding how computers can learn from data. Humans and animals can learn from the data that comes into their senses. For example, we learn to visually recognise objects and people, we learn to understand speech, drive cars, and play sports and games. Virtually all aspects of intelligent behaviour rely on some form of learning from data.
We can understand the basics of machine learning with an example. Take a task that humans might be good at: for example detecting people in an image. Getting computers to do this would be very useful for example when building a self-driving car that needs to detect pedestrians on crosswalks whilst driving. You can train a computer to detect people by giving it many thousands of images, with and without people in them, and labelling the ones with people in them. The computerised person detector is an algorithm (a computer program) that has many tuneable parameters. As it sees more labelled data, it tunes its parameters so as to improve its ability to detect people, minimising the number of errors it makes. The goal is to perform well enough that the person detector generalises correctly to new unseen data. Machine learning is different from typical computer programming because the program is not fixed; it changes over time as it experiences more data. In this way, machine learning is similar to human and animal learning; our brains also tune parameters (connections between neurons) in response to sensory data.
CSF: Could machines ever truly learn in the same way that humans do?
ZG: We don’t really understand how humans learn yet; truly understanding this requires integrating fields of research ranging from education to psychology to the molecular biology of individual cells in the brain. However, we don’t need machines to learn in the ‘same way’ as humans, just like we don’t need airplanes to fly in the same way as birds do. In fact, liberated from some of the constraints of biology (such as the size and energy limits on a single brain, the slow information transmission times of neurons, and other idiosyncrasies of evolution) we can imagine machine learning systems that can perform certain tasks far better than humans. We already have computers that can do (albeit imperfectly) many tasks better than I can do: for example, translating between languages, navigating in new cities suggesting routes that adapt to traffic, or analysing biomedical data to find genetic markers for diseases. We already have computers that can play many games (e.g. Chess, Jeopardy, Atari and Go) at expert human level.
Soon, we may have self-driving cars that rely on machine learning for many of their basic operations. The list of things that machines will eventually do better than we do will continue to increase.
CSF: What does this all mean for us and our future?
ZG: Machine learning systems are generally used as tools to make us more efficient. Like all technologies this will have an impact on the way we live; how we use these tools is up to us as a society. Ideally, we can channel these exciting technologies to provide for higher productivity, safer and more efficient transport, better and more personalised healthcare, and in general advancement in the pace of scientific discovery. These systems will also have an impact on employment, which we should probably address with safety nets and life-long training. Machine learning systems will have security applications (both positive and negative) that we need to manage.
CSF: How does mathematics come into it?
ZG: Machine learning is fundamentally a mathematical field. It’s not about tinkering with computer programs until they learn; advances have come from a fundamental understanding of the mathematics of learning systems. Again, to use the analogy with birds and airplanes, although airplanes do not need feathers and flapping wings to fly, understanding flight in both birds and airplanes requires basic physics, aerodynamics, and other concepts that are expressed mathematically. Similarly, there are equations for how to adapt the parameters of a learning system, how to combine information from multiple sources, and how to make optimal decisions. These equations are helpful both for understanding how machines could learn and how humans and animals learn.
CSF: What are the links to artificial intelligence?
ZG: Machine learning underlies many of the recent advances in artificial intelligence (AI), including robotics, computer vision and speech and language processing, and game-playing computers. Historically, machine learning was a reaction to traditional AI, which focused perhaps too much on symbolic logic and too little on statistical learning from data. The ML community divorced itself from traditional AI in the 1980s, but in the last few years the two communities have merged again.
In general, machine learning is a less controversial topic that relates immediately to many of the tools we use on a daily basis on, for example, our smartphones. On the other hand, AI is sometimes seen in a threatening way through the lens of science fiction—an unfortunate association that is not warranted by the current state of AI research.
CSF: Are there any risks involved in our increasing reliance on machine learning? What if things go wrong?
ZG: Like all computer systems, one should design safeguards to ensure that any damage can be controlled. There is an active research community trying to build more powerful safeguards so as to have transparency, trustworthiness, and robustness of such systems. I’m involved in the new Leverhulme Centre for the Future of Intelligence which is going to build a community of researchers around both short and long-term issues arising from machine learning and AI. The Royal Society also has an ongoing policy project on the impact of machine learning.
CSF: What are the positives? What has machine learning given us so far?
ZG: Machine learning is behind many, many of the things we now take for granted: the little rectangle that appears in our cameras detecting faces, email spam detectors, the ability of computers to translate text and recognise speech, systems that recommend items on Amazon and other online shopping sites, all online advertising, and all aspects of the user interface at Facebook. The data revolution has meant that there is a tremendous amount of data online, and machine learning systems thrive on this data.
CSF: What can we expect from machine learning in the next five years?
ZG: It’s hard to predict, but I would love to see it help revolutionise medicine for example, increase efficiency in transport systems, and in general improve the quality of life for people.