A Child Learns a Cat From Its Tail After a Few Glimpses. AI Needs Vast Image Libraries.

Show a toddler a few cats, then hide the next one behind a bookshelf so only its tail sticks out. The toddler will still say cat. A computer can now make the same guess, but only after training on vast amounts of images, and how the child does so much with so little remains a mystery researchers have not fully solved.

Both are making a bet, not stating a certainty. Fei-Fei Li, a computer scientist who studies artificial vision, explains that recognition is a probability judgment. A furry tail indoors is probably a cat because foxes, which also have furry tails, are rarely indoors.

Older computer programs tried to copy that logic with rules. Li says programmers told machines to first recognize the furniture, decide the scene was indoors, and then choose among only ten possible animals. Narrowing the choices made the guess easier.

Modern AI works differently. It learns from scale. After training on vast collections of photos and captions, the machine adjusts what researchers call learned weights, the strengths of connections inside the artificial network. When a new tail appears, those strengths tilt the machine toward the pattern it has seen most often, a cat tail.

A child does not get that much practice. Li points out that a child may have seen only three or ten cats in real life, yet still tells a cat tail from a fox tail. The child sees the cat from the side and from the back. The child sees only part of it while it moves. The child sees only a tail behind books. Li says the child learns "through a different kind of learning pathway." Scientists have not fully solved how that pathway works.

The next step is also about odds. Once the brain or the machine knows cat, it learns simple rules. The cat usually moves headfirst, not tailfirst. It might chase a mouse, and it might run from a dog.