Go exploring

Space is the stuff all around us – up, down, back, front, left, right. Wherever we go we all move through space. Now imagine instead that the space you inhabit isn’t filled with space but is filled with faces. Welcome to the strange world of face space.

Enter face space

We’ll try and give you an idea of what face space is like. Start by imagining a room. On one side of your room is an open door and on the other is a window. Starting in the middle, as you move towards the window things get lighter, and as you move toward the door you get more cooling air wafting in from outside. At any point in between the two you have a specific combination of light and breeze. Now let’s teleport the room to face space, so that on one side of this new ‘room’ is a really narrow nose, on the other side is a really broad nose. Now take any nose – how about yours if you’re not using it at the moment? You can place your nose at a specific position between the two extreme noses so that it’s just the right distance from both. Right in the middle would be the average nose, created by combining all the noses you know. Your nose might be wider than average or narrower than average, and you could let people know exactly what your nose looks like by giving them the average nose measurements and an indication of how much broader or narrower your nose was from this. Knowing the average and how far you are from it in each direction are the signposts in face space.

Your face looks like...

There are all sorts of ways to describe a face, like nose width, eye separation and so on. You can build up a complicated face space by deciding what the important features are, measuring lots of them, finding the average and then putting the bits making up individual faces at different positions relative to the average for those features. Eventually, researchers began to see evidence that there might actually be an average face stored in our brains too. We tend to recognise distinctive faces more easily. They are further from the average so we get a stronger signal. An average face shape doesn’t seem to carry any particular information about a person’s identity – it’s what’s different from the average that lets us identify people. By using a series of cartoons of varying levels of extremeness researchers showed that certain cells in the brain became more active the more different the face presented is from the average face. Somewhere deep inside our brain, it seems, we have learned an average face to compare other faces against.

Computers – our GPS in face space

To create this face space in a computer we first take lots of pictures of faces. We can align the features like the eyes and nose by changing the size and the orientation of the images. A picture of a face is just a massive number of pixels, where each pixel is the colour value needed at that particular point in the image – for example the shiny white of the teeth. With hundreds of face pictures all aligned, it’s easy to calculate the average face. Just add the values for all the images together, pixel by pixel, and divide the final value for each pixel by the number of images used.

But there is more information we can extract. We know about the average so let’s take that away from our information and look at what’s left. We can use a method called ‘principal components analysis’ to find a set of images that we can add to the average face to let us recreate any of the original faces. In effect we convert our huge combined face into a series of independent images, which then become the staging posts in face space, starting from the average face at the centre. So any face has a place in the space, a certain distance and direction away from the central average face. As we move out into the space, we combine the different component images to recreate the face from our original set that lies at our current position. In essence, we can determine positions in face space much like GPS can on the Earth.

Girl face vs boy face

That’s not all we can do. We can turn boys into girls! We can take a load of female faces, find the average and the staging post faces, and we can do the same with male faces. For any particular male face we can calculate its distance and direction from the average male face at the centre. Now, imagine a line on a map joining the starting place to the finish position. We can take the line we’ve drawn in the male face space and draw the same one in the female face space. Anchor its starting point at the average female face and see what female face it ends at. This will be the closest female equivalent of the male face. By moving in face space we can change the gender of faces.

Enter the antiface

We can also change the direction of our line in face space so it points backwards. If you follow this line all the way through the centre, then continue for the same distance out the other side, you’ll end up at the antiface: the face that is essentially the opposite of the original face. People rate these antifaces as being more dissimilar from the original than randomly created faces. It seems our brain knows its way around the face space we create in a computer, which might mean we keep a face space in our heads too.

Facing forward

You can probably see that recognising faces is a very complex job for our brains to do. Fortunately, it looks as though we have particular parts of our brains that specialise in helping us perceive faces. By building computer software that, as with face spaces, mimics the way our brains understand faces, we can start to explore and test our understanding of them. Then these discoveries can help us build new technologies, like robots that can recognise and express emotions (see pages 6 and 18). That should give us something to smile about in the future.