AI study gives insights into why super-recognisers excel at identifying faces

They have been used in the search for the Salisbury novichok poisoners, finding murder suspects and even spotting sexual predators. Now, research has revealed fresh insights into why super-recognisers are so good at identifying faces. Previous research has suggested people with an extraordinary ability to recognise people look at more areas across a face than typical people. Now researchers have used a type of AI to reveal how this approach aids their prowess. “It’s not just about looking everywhere – it’s about looking smart,” said Dr James Dunn, first author of the study from UNSW Sydney. Writing in the journal Proceedings of the Royal Society B: Biological Sciences, Dunn and colleagues report how they drew on eye-tracking data from one of their previous studies involving 37 super-recognisers and 68 typical recognisers. In that work, participants were shown both pictures of full faces and ones where the area of the face they were looking at was made partly visible. In the new study, the team used this data to reconstruct the actual visual information seen by participants’ eyes. This “retinal information” was then fed into deep neural networks (DNNs) – a type of AI system – that were trained to recognise faces. They also gave the AI system a full image of either the same face the participant had seen or a different face. In each case the AI system produced a score for how similar the retinal information was to the full facial image it had been given. The team compared the results of typical participants and super-recognisers as well as data based on randomly selected areas of the initial facial image. The results reveal in all cases the performance of the AI system increased as the parts of the face being looked at were made more visible. Also, across all levels of visibility, the performance of the AI system was highest when based on retinal information from super-recognisers. “It shows that differences in face recognition ability partly stem from how we actively explore and sample visual information, not only from later processing by the brain,” said Dunn. The team then investigated whether the results were simply down to super-recognisers looking at more areas of a face, and hence taking in more information. However they found that even when the amount of the face captured in the retinal information was the same, the AI system did better when fed data from super-recognisers. “This means their advantage isn’t just about quantity, it’s about quality,” said Dunn. “They pick regions that carry more identity clues, so each ‘pixel’ they choose is more valuable for recognising a face.” Dr Rachel Bennetts, an expert in face processing at Brunel University of London who was not involved in the work, welcomed the study. “To me, its main contribution to our understanding of super-recognition is the conclusion that superior face recognition isn’t just about looking a specific area, or looking longer or at more places on a face. Super-recognisers are exploring the face more broadly, but also sampling more useful information,” she said. Dr Alejandro Estudillo from Bournemouth University said the study was based on showing people still images under highly controlled conditions. “It will be important to test whether the same pattern holds in more naturalistic, dynamic scenarios,” he said. While the study suggests there are tactics that can aid facial recognition, it seems unlikely everyone can become a super-recogniser. “At the moment we don’t know whether these eye movement patterns could be trained effectively,” said Bennetts. Dunn said studies had suggested super-recognition was rooted in genetics and was heritable. “Super-recognisers seem to naturally pick out the most useful features, and that’s hard to train because it varies from face to face,” he said. The researchers have developed a free test to help identify super-recognisers which is available at UNSW Face Test.