UK Invests hard in Machine Learning and AI

The UK's leading science funding body, UKRI - UK Research and Innovation - announced a significant investment in artificial intelligence - or A.I. - based research. The purpose of the 200 million pound initiative which will be invested across the country is to help the UK to maintain its status as a world leader in this sector. Cambridge is one of the centres awarded funding where researchers are going to harness A.I. to enable them to sift through massive datasets looking for patterns that the human brain could never spot, in things like climate data and earthquake measurements. Chris Smith spoke to Scott Hosking,  a member of the initiative. Scott is at the British Antarctic Survey and he uses machine learning tools like this to understand climate change...

Scott - This is super exciting; this is an absolute game-changer for Cambridge and also the UK climate community. So our data sets are getting larger and larger year on year, and it's fantastic that we've got some extra help and these algorithms to help us sift through that data.

Chris - What will you spend it on? You’re going to get about six million Pounds worth of funding over, initially, five years with this aren't you? So how would you be managing the project? What are you going to spend the money on?

Scott - So this is a centre for doctoral training, so this is a five year project, which brings in 50 students over the five years, but hopefully will bring in more than that. We have all this industrial funding, so we have Google on board, and Microsoft, and there's over 30 partners in total. So this is big.

Chris - Okay so you're going to be effectively investing in the next generation if you're going for PhD students? These are early career researchers.

Scott - Absolutely. We need these new algorithms we need these new tools and we need to build that expertise first, so we are building the next generation of climate scientists that also have this machine learning artificial intelligence knowhow, which is something my generation didn't have.

Chris - So it's literally investing in the project; it's investing in technology and that sort of research and development, but also a strong investment in people?

Scott - A huge investment in people, exactly. And we'll benefit from that as the general public to improve our future climate predictions. We need these large datasets in order to look at extreme events. For instance, there's no point looking at a one in a thousand year event if you only have a thousand years’ worth of future climate data. So we really need to be running tens-, hundreds of thousands of years into the future and to do that we need fast climate models.

So one thing we're looking at is including AI / machine learning algorithms in the models themselves to speed them up. And also once we have all that data, how are we going to analyse it? As scientists, we really struggle to use traditional tools just to zoom in. If you're interested, for instance, in heatwaves in London, we may just zoom in over Europe but actually we should really be looking at all the data we have and look for those patterns in the data. Maybe there's something in Brazil, or a feature or something we've seen in the Arctic, which is very relevant to our climate, so we can feed all that in.

Chris - So - explaining to people just for a second how this actually works - when we're making predictions about what the climate is going to do in the future, is it fair to say you're essentially getting together enormous numbers of measurements - that might be temperature, it might be pressure measurements, it might be how wind speed is changing and pressures are changing on different parts of the Earth’s surface - I'm just speculating here - but then asking a human at the moment to try to spot patterns in all of that. Whereas, if you ask a computer to relentlessly go through and explore the relationships between all these numbers and all these enormous complexities, it will spot the needle in the haystack that we can't?

Scott - That's right. So the data not only is vast but it's also various. So we have all sorts of information - satellite data, climate model data, all with different variables, different weather variables say temperature humidity pressure etc. and just trying to picture that in anyone's head is just unfathomable. So we need those computer models which can build these really complex large matrices, multi-dimensional systems and search for those relationships, cross-validate things we may not even think is relevant. But actually if it does come out relevant that could be a game changer.

Chris - Now how does the A.I. or the machine learning side of it come into the equation?

Scott - So machine learning, sort of "under the hood", is just statistical algorithms that we've been using for decades. Learning is key here. All we're doing with these algorithms is looking for those relationships and providing an answer or a possible answer to a person. We're not doing AI at the moment because the I, the intelligence, suggests that we're going to do something with that data. Now in for self-driving cars the car needs to be intelligent to know whether to slam the brakes on. Our intelligence comes from the the businesses or the government officials that need to make those decisions. So the machine learning is that layer to provide decision makers with a robust set of tools, a robust set of analysis which they can make their decisions.

Chris - Researchers from many different fields are using these sorts of approaches now though aren't they? For instance in the last two years we've seen researchers take pictures of skin lesions and then ask a computer to learn what a healthy mole versus a potentially cancerous mole looks like. And by the time it's finished learning, it can outperform dermatologists who have been through umpteen years of medical school and board level exams to make sure that they're good doctors. So could this system effectively teach itself what to look for though?

Scott - Absolutely so these algorithms can do -  we can look at satellite images for instance, we could look at how disease spreads in forests, in vegetation, how different crops had their crop yields are suffering say to climate change and so these are things which the naked eye, are human eye, might struggle to pick out those signals, but a machine learning algorithm given enough data can see those signals.

Chris - And what does it then do with that information? Does it sort of flag to you and say “okay I've spotted this relationship, here's one that you need to now work on”?

Scott - Yeah. So we should never use machine learning as a black box and just trust the answer. You do need still maintain your expertise in climate science. Look at that information and say actually does this make sense? And maybe you'll go back to take a more traditional approach and build up a computer model to follow through a new theory.

Chris - I also mentioned earthquake data and things like that because I know that you're looking specifically what the climate has been doing will be doing. But you could take the same knowledge and the same approach, drawing huge amounts of information together to find out how this things and systems work, and apply it to many different things.

Scott - We can apply these algorithms all over the world so we can look at ice sheets and melting ice sheets and what that means for the communities in the Himalayas. There are 2 billion people that rely on this water. So these algorithms are an absolute game changer for people.