Wouldn't it be great if instead of sending humans into the hills to search for a terrorist, we could send an expendable mechanical device that would recognize his face and his voice and maybe even his personal mannerisms?
We're decades, if not centuries, away from that kind of capability, but scientists are making significant progress in developing machines that function in a way that is roughly similar to how humans think and sort out information. It's a really tough challenge because the human brain's neural network — our biological equivalent of a computer processor — is one fantastic piece of equipment.
Some of the best minds in the world are working in this arena, but how do you build a machine that can come even close to duplicating the human brain and perform tasks that are so simple we take them for granted?
"It's hard to do that with a machine," says Joel Davis, who manages cognitive, neural and bio-molecular research for the Office of Naval Research. But that's not going to keep him and others from trying.
Inspired by Biology
He believes the blueprints for building the amazing machines of the future that can figure out some things for themselves will come from the world of biology.
What biological systems have that even the most powerful computers lack, he says, "is something we call sensory fusion."
"You walk down the street and you're getting auditory input from your ears and visual input from your eyes and tactile input from your feet, but you don't see these as separate things," Davis says. "It's all fused together in one dynamic picture."
Those different streams of information arrive first in an area of the brain called the superior colliculus, and some amazing things happen there. The brain cells in that region react to more than one stimuli, and scientists have been able to measure changes in the neurons as they are stimulated.
"You can look at the cells that respond to both visual and auditory stimuli, and when you deliver both of these together you get a response that's more than the sum of both responses," Davis says. In other words, the brain doesn't just add the two responses. It multiplies them, greatly increasing the value of the data it's receiving.
"That's the biological neural network operating," Davis says. "There's some computation going on there."
In an attempt to transfer that capability to the world of machines, Davis is funding research at several universities, much of it aimed at replicating the remarkable function that we call vision. You can do that, it seems, by building a brainy camera.
The talk of the campus at the University of Illinois, Urbana, is a remarkable gizmo called SAC, which stands for self aiming camera. SAC knows how to pick out a person who's talking in a crowded room, and how to ignore other sights and noises that are occurring at the same time.
"It works pretty well," says Tom Anastasio, a computational neural scientist who is the principal investigator on the project. "Often it looks exactly at what you want it to look at."
SAC works on the principle that the superior colliculus does far more than simply receive data from sensory sources. It computes the probability that it ought to be looking at something, Anastasio says.
"SAC currently receives video and audio input from a conventional video camera and a pair of microphones," he adds. "It takes those two stimuli and it combines them, to compute the probability that a target has appeared in the environment. And then it initiates the movement of the camera to look in the direction of that target."
A Camera That Listens
The camera has been taught to distinguish between different sounds, so it knows if a person is speaking, or someone just dropped a book.
"The human voice has a very specific audio signature," Anastasio says. "It doesn't matter who's speaking, or what language, or the age or sex of the individual, they all seem to have a particular pattern in the audio power of their speech."
So SAC knows to zero in on a person who is speaking, rather than someone who is walking around the room, because it responds preferentially to the sound of the human voice. That could come in handy some day in automated video conferencing equipment.
SAC is very limited, however, in that it has to be taught just what to look for. The next goal, Anastasio says, is to have SAC learn for itself which targets in its environment are interesting.
"We don't want to have preconceived notions of what targets to look for," he says. "We want to detect a source of new information in our environment and turn around and look at it and see what it is. It might have survival value for us. Maybe it's a danger thing."
That's a tall order, but there are more modest applications for this kind of research, and those partly explain the military's interest. The Office of Naval Research's Davis, for example, thinks an advanced version of SAC might be useful in "directing counter fire."
Armed and Robotic
A brainy camera, armed with microphones, could train a gun automatically to return the fire from a hostile source.
The camera could pick up muzzle flashes, and the microphones could detect sounds consistent with gunfire, and almost instantaneously fire back.
And the sensors need not be limited to the range of human capability. Infrared sensors, for instance, could detect the presence of a living organism, radar could detect the distance to the target, and other sensors might even be able to sniff out dangerous fumes. Then it would combine those separate streams of data and compute whether the target poses any danger.
Of course, you couldn't call this thing SAC.