Synthetic Learning

Annual Question: 


This year, researchers at DeepMind, an AI company in London, reported that they taught a computer system how to learn to play 49 simple video games. They didn't teach it "how to play video games," but how to learn to play the games. This is a profound difference. Playing a video game, even one as simple as the 1970s classic game Pong, requires a suite of sophisticated perception, anticipation, and cognitive skills. A dozen years ago, no algorithms could perform these tasks, but today these game-playing codes are embedded in most computer games. When you play a 2015 video game you are usually playing against refined algorithms crafted by genius human coders. But rather than program this new set of algorithms to play a game, the DeepMind AI team programmed their machine to learn how to play the games. The algorithm (a deep neural network) started out with no success in the game and no skill or strategy and then assembled its own code as it played the game, by being rewarded for improving. The technical term is unsupervised learning. By the end of hundreds of rounds, the neural net could play the game as well as human players, sometimes better. In every sense of the word, it learned how to play the games.

This learning should not be equated with "human intelligence." The mechanics of its learning are vastly different from how we learn. It is not going to displace humans or take over the world. However, this kind of synthetic learning will grow in capabilities. The significant news is that learning—real unsupervised learning—can be synthesized. Once learning can be synthesized it can be distributed into all kinds of ordinary devices and functions. It can be used to enable self-driving cars to get better, or for medical diagnosing programs to improve with use.

Learning, like many other attributes we thought only humans owned, turns out to be something we can program machines to do. Learning can be automated. While simple second-order learning (learning how to learn) was once rare and precious, it will now become routine and common. Just like tireless powerful motors, and speedy communications a century ago, learning will quickly become the norm in our built world.  All kinds of simple things will learn. Automated synthetic learning won't make your oven as smart as you, but it will make better bread.

Very soon smart things won't be enough. Now that we know how to synthesize learning, we'll expect all things to automatically improve as they are used, just as DeepMind's game learner did.  Our surprise in years to come will be in the many unlikely places we'll be able to implant synthetic learning.

Date: 

[ Thu. Dec. 31. 2015 ]