Computers and humans learn in very different ways. Humans see one example and can intuitively figure out what an object or symbol might mean. Computers, however, have to be fed thousands of examples before they can do the same.
Researchers from the New York University have developed a way to mimic how humans take mental leaps when they're learning. Not only that, but the machine can also recreate simple symbols and drawings in a way that is extremely similar to those made by humans.
The research was published in Science and describes the creation of a "Bayesian Program Learning (BPL)," which essentially turns a concept into a computer program, allowing computers to learn from a single example.
The model is also able to use knowledge from previous concepts to learn how something works. For example, if a computer already knows the Latin alphabet, it can much more easily learn the Greek alphabet, which is similar.
"But what the program learns - its concepts - are also programs. We think that is true for humans too: your concepts are programs, or parts of programs," said Joshua Tenenbaum from MIT in a presentation.
Perhaps most notable is the fact that when a computer was told to create new examples based on the concept that it learned, the images were extremely similar to examples that humans had created. In fact, other humans couldn't distinguish whether a person or computer had created specific examples.
The research could have significant implications. It could help computers learn new things much quicker, enabling robots and computers to adapt to new situations and scenarios without having to be fed thousands of previous examples.
Of course, applications for the research are years away. The researchers themselves note that the system doesn't see anywhere near the same level of details as humans, and that it lacks knowledge of structural things like parallel lines and optional elements to symbols, such as a cross bar in the number seven.
