The Human CPU: from Computer Science to Neuroscience.

Chris - We're joined by Demis Hassabis, who traded in a career designing and programming computer games such as Theme Park to become a neuroscientist! Tell us your story. You started off as a computer science student.

Demis - That's right. So my undergrad degree was in Computer Science at Cambridge and I was having a traditional root in computer science then. When I left I set up my own computer games company which I'd also programmed some games like Theme Park before coming up to Cambridge. It was a natural progression for me to do that. I set up my own company in Camden in London and we grew it up to 65-70 people and created several games for big publishers around the world. And then about 3-4 years ago I decided that the games industry was going in a slightly different direction to the way I thought it would do when I first got into games in the early nineties. It was becoming a more big business, big budget sort of Hollywood-type industry where there was less and less room for creativity, I was feeling.

Chris - Do you think the financiers noticed the scale of this market? Previously it was a bit geeky, bit niche, not a very big market - still cost quite a bit. Because obviously, you're saying big teams of people were needed to do this but the profits weren't huge so no one was that interested. Suddenly people realised you can tether a computer game to a Hollywood blockbuster, you can make a fortune.

Demis - Yeah, I think partly the realisation how big a business it was but I think the main problem and the reason that caused the conservatism, if you like, in terms of creative sense was that games became very expensive to make. The graphics got so sophisticated you needed teams of - the TripleA game now would have easily 50 artists on it. The costs of that are just huge. Because it is costing 10 million pounds for one game the money men at these publishers would need to be more sure there was definitely going to be a return. This means linking it in with a Hollywood franchise.

Chris - That obviously was inconsistently compatible with your view of what you wanted to do. So you did the rather unusual thing of take a side-step into neuroscience.

Demis - Yeah, that's right. Although I had been in the games industry for a long while, underlying all that all the games I'd been involved with designing and programming actually composed a lot of AI in those games. Most of the games were big strategy

Chris - Artificial Intelligence for the non-initiated like me.

Demis - That's right. All the games I wrote like Theme Park and Republic and Evil Genius. They all involved simulations. Most of the involved hundred of little computer people coming in and getting involved with the game environment. Most of the games, like Theme Park, involved you manipulating that environment and seeing how these autonomous agents reacted to what you were doing. Those were the kinds of games I found fun to play and they were the kinds of games I found fun to create. But so underlying all of this was my passion. My main passion is actually in artificial intelligence and related to that (as soon as you start thinking about what artificial intelligence is) then you start thinking about - how is it the mind achieves these end-goals?

Chris - You didn't get hooked on your own games, did you?

Demis - No. Actually after you've worked on a game for 3-4 years you're sick of the sight of it in general, even if it's the best game ever!

Chris - It used to take me about ten minutes to get bored of them. Some of the strategy games I think are terrific because they forced you to think in a certain way. I was a very big fan of text games in the early days. Largely computers were rubbish and the only thing they could do was to generate spurious streams of text that you could read laboriously. They taught me amazing language skills. I think it's partly responsible for why I have an incredible memory for text and words and facts. You had to remember effectively a graphical representation in your head of the text environment that was this computer adventure. I think that probably had a huge brain training effect on me.

Demis - Absolutely. I think those early games left a lot more to your imagination in the way a book would, a great book. It exercises you imagination which, of course, with all the latest flashy graphics although it looks very beautiful obviously leaves less need for visualisation and creative powers.

Chris - Now just very briefly, how did you develop an interest in neuroscience? How did you then take those skills that you had in computer programming to start answering important questions about how the brain works?

Demis - I didn't really know much about neuroscience before I did my PhD but I did a lot of reading and it struck me that a computational approach or an understanding of algorithms or the basic computer science that Professor Bishop was talking about is actually a useful approach maybe to take looking at how the brain works. In general most of the people work with and most people are from a life sciences or medical background which, of course, is incredibly useful form an anatomical point of view. There are actually relatively few people in the neurosciences field that look at the brain in a computational way, a machine-like way which can give other insights. It's not a machine, the brain, but a lot of the things it does are relatively machine-like.

Chris - And the problem you solved most recently in a nutshell, what was that?

Demis - My most recent study involved trying to accurately predict where someone was standing in a virtual reality room whilst they were lying in a scanner, just from their brain scans.

Chris - Why is that so difficult a problem to solve? I thought there were cells in the brain that fire off when you go into a certain place. We can just read what that activity is, can't we?

Demis - it's been known for a good 30 years that these experiments were done in rats. There are cells that tell you where a rat is positioned in an environment or a pen. No one really knew what those cells looked like on a population level. If you were to look at a million of those cells at once, which you can't do with a single cell recording from electrodes in rats' brains. What I wondered was if you could have this global eagle's eye view of the whole population of cells that might tell you something different that you couldn't see on the individual cell level.

克里斯-这是应该是一个模型或reprentation of what's going on in the real world? Are you trying to understand how, because the person you worked with, Eleanor Maguire very famously worked out how our cab drivers have bigger bits of their brains through navigating round London. Are you trying to basically understand how a cab driver finds his way around London?

黛米-我们试图理解这些结果,really. We know the hippocampus is vital for spatial navigation and spatial memories but we still don't know what it is about the environment that is encoded in those memories and is encoded by those neurons and how and why it's doing that. You can't carry a two tonne brain scanner round with you. The best thing you can do is to record people whilst they're navigating round a virtual, very realistic virtual reality environment.

Chris - When you do this, what did you see and how does this influence our understanding of how the brain tells us how to get from A to B?

Demis - Brain scans work at quite a low resolution compared to single cell recording. When you look at one pixel, if you like, of a brain scan which is called a voxel - a 3D pixel. That contains about 40,000 neurons. For our pattern recognition algorithms you need to classify which location that was related to. It needs several hundred voxels to be active to train a pattern recognition algorithm with. What that shows is the spatial memories must be represented by very large neural populations' codes. Including probably between two and five million neurons.

Chris - Where previously we thought it was just the odd cell that fired off. Now you're saying there's huge populations of cells that join together. They code specific locations.

Demis - That's right. Also it says something about the way that they're clustered and structured. We found that these cells were also clustered quite closely together whereas previous rat literature suggested that perhaps they were random and uniformly distributed across the hippocampus.