A huge breakthrough from researchers at NYU could mean that computers are about to surpass humans in their ability to learn.
Artificial intelligence has been a prominent theme in science fiction works throughout the 20th century, but only in recent years has the technology made significant strides towards becoming a reality. According to a report from Mashable, researchers from New York University have made a huge breakthrough in teaching computers to learn much like humans do.
Computers could reach conclusions after sifting through hundreds of thousands of examples, but scientists have had a hard time getting them to make inferences and comparisons during machine learning. The new study, however, shows how computers can create symbols and drawings that are nearly indistinguishable from those made by humans.
The study, published in the journal Science, describes how scientists developed a “Bayesian Program Learning,” or BPL algorithm. This allows a computer to turn concepts into simple programs, giving them the ability to “learn” large sets of visual concepts from a single image.
For example, the computer could generate the code needed to display the letter “a” and its variations, so that it would be able to read and recreate handwriting. The computer can build off of its previous knowledge to learn new characters and images, which can be extremely useful for translation purposes, for example.
According to Moore-Sloan Data Science Fellow at NYU, Brenden Lake, “If you ask a handful of people to draw a novel character, there is remarkable consistency n the way people draw… They do not see characters as just static visual objects. Instead people see richer structure… that describes how to efficiently produce new examples of the concept. We aimed to develop an algorithm with the same capability and then compare it with people.”
The computer was capable of creating new examples based on an original concept, many of which were created by humans. The end result was nearly indistinguishable from a person’s work. The computer was able to pass a rough iteration of the Turing Test – which says that at a certain point a computer would be 70 percent likely to convince a person they were talking to someone else.
A press release from NYU outlining the study’s findings can be read here.
