• University of Glasgow experts investigated how we remember each other’s faces
• Researchers had people compare their memory of faces with similar 3D models
• From this they could work out which features our brains use to identify faces
• The findings could have applications to AI, gaming and eyewitness testimony

It’s great to spot a familiar face — and now researchers have ‘cracked the code’ that our brains use to tell one apart from another.

Our memories of the faces of people we know focus on key facial features let us recognise them when we meet.

Volunteers were asked to rank how closely randomly-created digital faces matched with their memory of the face of a colleague.

This process was repeated over and over, revealing key identifying facial features of the colleague being remembered.

Computer software analysed the data on these instances of key features to recreate the faces in question. 

The findings could be used to reconstruct faces described by witnesses to a crime more accurately, as well improvements in gaming and artificial intelligence. 

Researchers from the University of Glasgow showed pictures of four of their colleagues’ faces to fourteen other university members.

They then set about trying to determine which specific facial features that the participants used to identify their colleagues’ faces from memory.

Volunteers repeatedly compared their recollection of one of the four real people’s faces with six randomly generated faces of the same age, ethnicity and gender.

Participants then picked the randomly generated face that was most like their memory of the real person’s face and ranked how similar they were.

Researchers were then able to work backwards to determine which physical features are relied on to remember a given face.

‘It’s difficult to understand what information people store in their memory when they recognise familiar faces,’ said paper author and psychologist Philippe Schyns.

‘But we have developed a tool which has essentially given us a method to do just that.

‘By reverse engineering the information that characterises someone’s identity, and then mathematically representing it, we were then able to render it graphically.’ 

This let experts build up a picture of which features were key in our recollection of different faces — which they used to create a mathematical face-producing model (stock image)

Researchers then used a database of 355 digitally-captured faces, each of which was characterised by its shape and texture, to create a generative model of 3D face identity. 

The team then used this to model to test the validity of their results, asking a new set of participants to rate the similarity between their recollections of a familiar face with random faces generated by the model.

By keeping the same shape and texture information relating to age, ethnicity and sex as the real faces, the researchers could isolate each face’s unique identity information.

This information was then used to produce entirely new faces. This stage of the process could be used to generate hyper-realistic faces for computer game characters, for example.

The full findings of the study were published in the journal Nature Human Behaviour.