What is your thesis about?

‘The thread running through my thesis is ‘prediction’. This is not something we learn in our studies. We are taught to think in terms of hypotheses, to experiment to see if there’s a difference between condition A and condition B. But I never felt this was the right way to be finding answers about the brain – we can keep doing test after test and study after study looking at individual elements, but my assertion is that won’t bring us to the truth about how we think, how we behave. I believe we should be looking at what the power of artificial intelligence can do in this area, and that’s why I decided to use machine learning to do something radically different during my PhD.’

What was new about what you did?

‘As psychologists and neuroscientists we’ve been taught to think in ways where you tightly control everything - one input variable, one output variable - so the only thing we can conclude from anything we do is the relationship between the two things we measure. But when it comes to the human brain these represent only tiny pieces of the puzzle.

‘I feel this is a fundamental difference between the newer sciences and the older ones - in physics there are certain laws which are unchanging and you can therefore use physics to land a spaceship on the moon, for example, because you know exactly how the forces out there in space are going to act on your spaceship. But there are no simple laws governing human emotions. Testing individual reactions in isolation seems to miss the point when the brain is one of the most complex things we can imagine. And my feeling is that we should lean into that: let’s build more complex models, with more data, more variables, faster computers to see to what extent we can predict human behaviour.’

How did this manifest itself in your experiments?

‘My thesis encompasses quite a few different experiments. In one, for example, we wanted to look at which elements of another’s face we use when decoding their emotional expressions. Instead of taking the standard approach and investigating the effect of just one element - say, whether the corners of the mouth are turned up - we used computers to generate faces according to random combinations of 20 different facial movements. Then we made a model that would predict what a participant would see in a given face with its combination of movements.

Then we went one step further. We know people use facial movement when inferring emotions, but a lot of research shows we’re biased by the face itself, so the structure regardless of the movement, what we call ‘facial morphology’. For example, men with broad faces are more easily interpreted as being angry, regardless of facial movements. So then we added random morphological features – longer faces, slightly more wrinkly faces, bigger nose, eyes closer together -  to our faces with random movements. We wanted to know if our model would it improve, would it predict outcomes better? And it did. This a very complex model, with 100s of variables, and it isn’t something we’d have been able to demonstrate if we’d used a simple model.

I not particularly interested in how much each variable of the experiment is related to the outcome, but I’m very interested in the outcome as a whole, and how adding more elements actually made this model better. It’s way more complex, but the extra accuracy in prediction we get shows us, I think, that it’s doing the same sort of thing that humans do. In the same way the human brain comes up with one output – that person is happy or sad, for example – but goes through an amazing amount of computation to get there, and that’s what this model is doing too.’

So the model is successfully simulating a part of the human brain?

‘Of course, then you run into what’s known as the ‘black box’ problem of AI. We can build these super complex models, trained used millions of different data points and billions of parameters, and these models predict extremely well. But the thing is, we don’t know what’s going on inside. Just like we don’t understand the brain, we also don’t understand the model that models the brain. But I’d rather have that black box modelling the brain successfully, because then I can’t start tinkering with it, looking inside it. That’s where my research will head in the future, trying to build ever more accurate models that reflect the human brain and human behaviour in all their complexity.’

Defence details
Lukas Snoek: Towards prediction: Studying the mind and brain in the age of machine learning. Supervisors are Dr H. Steven Scholte and Dr Suzanne Oosterwijk.

Time and location
Agnietenkapel on Wednesday, 9 February, at 1 p.m.