For a man who would grow up to become a leading expert on teaching computers to think, Peter Tu spent his childhood on some unlikely hobbies. Rather than holing up in his bedroom to code, the young Tu would often run around his Toronto neighborhood with his dog or oversee the reptile pit he built in his backyard. Of particular interest to the future computer scientist were the ways his menagerie of animals learned new tricks and skills.
These days the 53-year-old Tu, chief scientist for artificial intelligence at GE Research, is using his eclectic interests to find new ways of training computers to learn — ways that mimic how humans and animals come to understand the world. When Tu talks about his research, he peppers his descriptions with phrases outside the vocabulary of ordinary computer scientists. Tu imagines artificial intelligence moving beyond simply analyzing huge data sets and toward learning to “get the gist” of a problem or extract meaning from “visceral experiences.”
While many humans make a habit of anthropomorphizing the machines they interact with, Tu’s colorful phrases are chosen carefully, offering hints of the revolutionary ideas that his team at GE Research is exploring in a collaboration with researchers at Siena College. Under a grant funded by the Defense Advanced Research Projects Agency’s Grounded Artificial Intelligence Language Acquisition (GAILA) program, this researcher is working to teach computers to learn language through visual and contextual cues — the same sort of cues that animals and children use to interpret the world.
Current AI systems excel at digesting massive sets of data and then making statistical inferences, thereby spotting hidden patterns too subtle for any human mind to recognize. That’s useful for a limited set of applications, such as analyzing billions of stock market transactions to spot new ways to make profitable stock trades. For other problems, which suffer from what Tu calls “a poverty of stimulus” — basically, not enough information — the standard bulk data analysis is useless.
Fortunately, he adds, getting information from limited data sets is just what children and animals excel at doing: “As children we do a large number of inductions. A child with limited examples is very good at drawing out general rules.”
The trick for children and computers, Tu explains, is knowing how general a hypothesis can be derived from limited examples. Extrapolating rules from a small set of experiences can go badly wrong when a human or a computer system overgeneralizes by imagining causal relationships where none exist.
To limit that problem, Tu is experimenting with providing young AI systems with the equivalent of artificial parents — experts that can step in and offer corrections. This is particularly useful in helping systems learn exceptions to seemingly general rules, such as the principle of English grammar that says you can describe a noun by putting its color ahead of it. That works for an orange shirt or a blue car, Tu points out, “but I wouldn’t say an ‘orange orange’ or a ‘blue wind.’ That breaks the rules.”
If Tu can figure out how AI-enabled machines can master such nuances in communicating with humans, and eventually with each other, he thinks it will be a natural step toward helping all sorts of machines comprehend all sorts of problems. His point is not that machines are taking the most logical route to the answer. Rather, the idea is simply that they could share their collective experiences to slowly evolve new solutions by trial and (many, many) errors, testing out and discarding metaphors in search of a computational model that achieves their goals.
The trick, says Tu, will be learning to accept the imprecision that comes from freeing the computer from fixed methods of tackling problems. “What we want to avoid is hardwiring intelligence into agents,” he says.
As for the larger questions of how close this sort of machine learning could one day come to approximating human intelligence, Tu says that it will progress far more slowly than the growth rate of basic computing power. Modern human brains have been evolving for some 10,000 generations, he points out, which provides a giant head start on the computers that will have to acquire knowledge from their own time-consuming experiments in the real world.
At heart an optimist, Tu says one of the things he loved about his childhood pets, especially his dog, was the idea that two species could evolve compatible forms of intelligence for their mutual benefit: “I’ve always been of the opinion that we need more companions to walk down this path with us.”