Thursday 26 February 2015

positronic-reinforcement

The New Yorker has a nice, succinct piece on the recent demonstration of the artificial intelligence DeepMind, whose talents draw from two sources, a deductive network of filters and positive-reinforcement.
The program—instructed with only the protocol that winning was good and losing bad—dazzled the human audience with a stellar progression on a platform of classic arcade games with some very masterful and unexpected strokes. It is not that DeepMind is inside the game, like when one challenges the game, but separated like a human player, and quickly devised a sure strategy. The program, however, did not perform quite so well with certain games—like Ms. Pac-Man, and the handlers weren’t quite sure why. Some disparaging voices checked their enthusiasm, as milestones like Deep Blue beating a chess grand-master or Watson winning against Jeopardy! quiz-masters. These achievements, though not coddled and not insignificant, came about, however, through extensive coaching, whereas DeepMind is learning on its own. What do you think? Is growth going to be exponential and get very quickly out of human hands?