This weekend, robot cars competed in a challenge that most humans would find trivial: drive 132 miles in 12 hours without crashing. Yet crash, they do. The difficult part isn’t so much the steering and acceleration, it’s determining the difference between an obstacle you must navigate around and a benign shadow on the road; it’s deciding whether that dark patch ahead is open roadway or deep water. These things are so easy for humans that we take them for granted, yet for a machine it’s a task literally in its infancy.
By the time a child is 2, it can easily tell the difference between a shadow and a real object, walk through an environment crowded with obstacles, do all the things that robots find most difficult. So how do babies learn these critical perceptual tasks that are so hard for computers? A team led by Paul Quinn investigated one aspect of this problem—how babies group similar objects.
Grouping objects is one way we are easily able to navigate through an environment: why, for example, do we see hundreds of leaves on a tree and recognize they are part of a solid object, but at the same time understand that the dapples of sunlight they filter onto the ground are not? One reason is that the leaves are all similar in shape, but the light admitted through the spaces between the leaves is not. You’ve probably heard the term “Gestalt,” which refers to the set of principles we use to make visual sense of the world. The principles are simple: similarly shaped objects should be grouped together, items moving together are probably part of the same object, things closer together are likely to be connected. Somehow, however, while humans and animals are great at putting these rules together in order to function successfully, the challenge of programming a computer to do the same thing is a daunting task.
What Quinn’s team wanted to study is whether babies are born with all of the Gestalt rules intact, or whether some of them are learned or acquired as they grow. A different team led by Quinn had established that babies as young as three months old are able to group objects that are similar in brightness or darkness. Now they wanted to examine whether babies can also group objects based on shape.
So how do you test such a thing in a baby too young to speak? Quinn et al. used a familiarization procedure: you show a baby one item until she’s bored with it. Then show her a two new items—one is the same, and one is different—and see how which one seems to hold her interest more. In their first experiment, they used arrays of Xs and Os:
The researchers reasoned that if babies could successfully group the shapes, they would be more interested in a different pattern of blocks than a similar one. Two groups of babies were tested: 3- to 4-month-olds, and 6- to 7-month-olds. The younger babies stared at both patterns of blocks for an equal amount of time. Older babies, however, looked at the different pattern (horizontal blocks after being familiarized to vertical rows, or vertical blocks after horizontal rows) significantly more of the time—57.58 percent, compared to 42.42 percent for the similar pattern.
But what if babies weren’t actually grouping the objects, but rather simply observing that the overall pattern had rotated? To address this concern, the team conducted a new experiment. In the first part, the procedure was the same, except that now babies were tested on patterns of Xs and Os instead of blocks:
But for the second part, babies were shown random patterns of Xs and Os:
The 3- to 4-month-olds again did not prefer either pattern, whether organized in rows or columns, or random. However, the 6- to 7-month-olds, as in the first experiment, preferred the rotated pattern of columns and rows (59.76 percent to 40.24 percent), but were ambivalent about the rotated random patterns, suggesting that the difference between the babies is really a difference in grouping ability.
But perhaps the problem wasn’t that 3- to 4-month-olds couldn’t group by shape, but that they couldn’t distinguish between Xs and Os. So they gave one final test:
Now both younger babies (62.95 to 37.05 percent) and older babies (66.50 to 33.50 percent) preferred the solid pattern to the identical pattern. Clearly, both groups of babies can tell the difference between an X and an O, so the evidence really does seem to indicate that older babies have learned to group by shape.
When the Gestalt principles were being uncovered in the 1920s, researchers argued that the rules were innate. But Quinn and his colleagues argue that their experiment shows that babies actually learn how to group objects based on shape sometime between 3 and 6 months of age—so it’s not an innate ability. With dozens of other principles to learn, and with an even more complex array of rules governing how the principles are applied, it’s astounding to realize that babies acquire them so quickly, while scientists struggle to duplicate the same tasks.
Eventually robots will be able to drive trucks through the desert, but it will have taken much more than two years to accomplish—and hundreds of other human problems, such as understanding language, creating art, and feeling true love, will remain unsolved.
Quinn, P.C., Bhatt, R.S., Brush, D., Grimes, A., & Sharpnack, H. (2002). Development of form similarity as a Gestalt grouping principle in infancy. Psychological Science, 13(4), 320-328.