Humans continue to be better at facial recognition than machines – though the incredible ability to process minute visual data continues to perplex scientists.
A team at Carnegie Mellon University is now looking at the millisecond-by-millisecond process of recognition through brain scans – and comparing the neural data to artificial intelligence systems.
Their work, which could conceivably help forensic science in eyewitness testimony, is published in the latest Proceedings of the National Academy of Sciences.
“Our results provide a step toward understanding the stages of information processing that begin when an image of a face first enters a person’s eye and unfold over the next few hundred milliseconds, until the person is able to recognize the identity of the face,” said Mark D. Vida, postdoctoral researcher at Carnegie Mellon’s Center for the Neural Basis of Cognition, and one of the authors.
Four people were showed pictures of 91 different people, with two different facial expressions, happy and neutral. The subjects reported when they recognized repetition of faces.
Throughout, their brains were scanned by magnetoencephalography, or MEG. The brain activity, from sight to the cognitive process of recognition, was cataloged by successive milliseconds in brain activity, they said.
Between 50 and 400 milliseconds, the two lightning-quick responses are produced by the right lateral occipital cortex, and the right fusiform gyrus.
After about 200 milliseconds, the viewer transitions from image recognition, to an assessment based on identity, they added.
The establishment of information patterns in the neurons – and potential feedback through later memories – could be better understood through further study, the authors contend.
“These results have implications for understanding the microgenesis of fine-grained, high-level neural representations of object identity, a process critical to human visual expertise, and perhaps for distinguishing between feedforward versus recurrent-feedback accounts of visual processing,” they write.
The neural data were also then compared to a computer simulation of an artificial neural network that recognizes faces.
“Combining the detailed timing information from MEG imaging with computational model of how the visual system work s as the potential to provide insight into the real-time brain processes underlying many other abilities beyond face recognition,” said David Plaut, a Carnegie Mellon professor of psychology, who was another of the authors.
Facial recognition systems have been improving recently, experts have said. However, the difficulty in recognizing faces and matching identities becomes more difficult once faces go from face-on angles toward profile angles.
Eyewitness accuracy has been a major contention in forensic science for several decades. Highly-publicized exonerations based on DNA have overturned convictions based largely on eyewitness testimony. Science is continuing to try and understand better how people can mistakenly identify other people. “Memory contamination” based on police interviews were the reason for uncertain eyewitnesses later sticking to their picking out someone of a lineup, according to a study of hundreds of Houston Police Department robbery cases published in the Proceedings of the National Academy of Sciences last December. A Dutch team found earlier this year that solitary eyewitnesses are not as accurate or reliable as couples.