To figure this out, scientists first showed monkeys a set of
pictures and recorded which face cells —
neurons that specifically respond to faces — fired and
which didn’t. What they found was that single cells weren’t
responding to single faces. Instead, each cell was encoding a
vector, or one direction in facial space, which means that a
single neuron may respond only to a certain distance between a
person’s eyes or a dimple on the left side of the mouth.
So, while there are tens of thousands of neurons tuned to
facial features, you only need around 200 to uniquely encode
any given face. The signals from just that small number of
cells can encode enough information to be able to differentiate
a new face from all of the others the brain has encountered.
Once they knew what facial characteristic each cell was
responding to, the researchers were then able to show a monkey
a new face, record the cellular activity in response to it and
reconstruct the face based on what the neurons were doing.
And they were able to do those reconstructions with remarkable
They were also able to create faces that looked completely
different from each other but as long as those faces shared one
specific feature, the researchers could get the same neuron to
fire in response to all of them. So, if entirely different
faces had the same crooked hairline, for example, the same
neuron would fire in response to every face.
While these findings still need to be replicated, this work
could help inform
technologies and AI. Le Chang, one of the researchers on
the project said in a
press release, “One can also imagine applications in
forensics where one could reconstruct the face of a criminal by
analyzing a witness’s brain activity.”
this video, to hear how a single neuron responds to a set