Monday, November 16, 2015

Teaching Machines to Learn on Their Own

Scientific American (11/10/15) Larry Greenemeier; Steve Mirsky 

In an interview, Xerox Palo Alto Research Center CEO Stephen Hoover discusses the swift changes machine learning is undergoing. He says computers' growing ability "to understand in much deeper ways what it is that we're asking and trying to do" is starting to be incorporated into products, such as the Nest thermostat. Nest, for example, has a built-in agent that learns from user behavior and infers context so it can anticipate how to operate. Hoover says machine learning involves the machine deducing the right answer from data input and programming itself, instead of the programmer breaking down a task into a series of steps. "You're going to show the computer a bunch of instances and you're going to label it, and it's going to learn how to do it," he says. "There's a core code which is that learning algorithm, and then that's applied to multiple contexts.' Hoover credits Moore's Law with enabling continued advances in machine learning. "Hardware not only begets the capability to create new kinds of software like machine learning, but also is creating new ways to sense, measure, and control the world," he says. "And that feedback loop is again one of the big changes that we're going to see coming."