Visual perception and baby chicks
Computers can do a lot of things faster and more accurately than humans. But there are a few areas where humans have machines beat. One of these is object recognition. It’s easy for people to recognize a cup or a dog or a car when they see it, but computers struggle.
Why is this such a hard problem for machines? Because objects don’t have a single set of visual features and configurations that define them. A cup looks different depending on how far away you are, the angle at which you’re looking, and the context in which it appears. Yet humans process this information effortlessly most of the time.
How is it that our brains represent the information associated with an object, despite the enormous variability in the actual perceptual information received across different instances? This is a core question in the field of vision science. One important piece of the puzzle requires understanding how the visual system is set up from birth.
How much does the visual system develop based on experience versus ability at birth? How critical is experience in viewing and interacting with objects when it comes to recognizing them from different perspectives? To study this using rigorous experimental methods, we would need to be able to completely control the perceptual experiences and visual environment of a baby from birth. Obviously, this is not a feasible or ethical thing to do in humans.
One alternative approach that some scientists have taken is to study chicks — infant chickens — which turn out to have similar visual processing systems to humans, even more so than many mammals.
Research from the lab of psychologist Justin Wood at the University of Southern California, supported by NSF’s Social, Behavioral and Economic Sciences Directorate, has resulted in some remarkable discoveries suggesting that, despite the complexities involved, the visual system seems able to handle object recognition from birth, with virtually no prior past experience! Wood takes advantage of the fact that chicks “imprint,” or become attached to objects seen shortly after birth, which is the way they learn to identify their mothers. By allowing chicks to imprint on particular objects, he can test whether they recognize those objects from different angles and in different contexts.
So, what have we learned from studying chicks’ ability to recognize their “attachment objects”?
Newborn chicks are able to visually recognize objects in spite of viewpoint variability, even when viewing the very first object they’ve ever seen and even when they’ve only seen it from a few different angles. This indicates that the visual system doesn’t need practice to recognize objects from multiple angles before it can accommodate changes in perspective.
Chicks remember the original views of the objects that they see but also can accommodate to new angles and contexts. This flexibility and sophistication of the visual system from birth is unexpected but important because it changes the kinds of questions we need to ask about how the visual system develops and processes objects. Although many aspects of the visual system require training, the heaviest lifting with respect to object recognition seems to have been solved through evolution.
This makes the challenge for computer vision all the more difficult at a time when our reliance on such technology is more important than ever. As challenges faced by the developers of driverless cars demonstrate, until computers can distinguish the side-on view of a tractor trailer from the sky, humans (and chicks) will have them beat hands-down every time.