A machine that learns by doing (and redoing)

- When I was in high school, I flew to London. To help pass the time, my parents bought me the Kasparov Electronic Chess Master. The game was barely a computer. It boasted 16 kilobytes of memory, so that was just enough to hold about a single page of text. On the side of the box, it said Your Intelligent Chess Companion. I played about a dozen times and each time it beat me. Nine hours into the flight, the gravely voiced person next to me leaned over and said, "I guess that thing's smarter than you are." As I packed up the box, I started to wonder if he was right. How could I have gotten beaten by this tiny $45 plastic box? After all, a mathematician once said that there are more variations in chess than there are atoms in the observable universe. That's way more than can fit into 16 kilobytes of memory. As it turns out, the machine wasn't necessarily smarter than I was. It was just a whole lot faster. Computer programmers had hand coded simple IF-THEN statements based on Boolean logic. Each time I made a move, the machine would speed through its logic. If the knight moves here, then move the rook here. Each of these IF-THEN statements was pre-programmed, all played out with the help of the man on the side of the box, the Chess Master, Gary Kasparov. This Boolean technique was good enough to beat me, but it wasn't creative. Remember, you still needed a chess master to program the logic. Kasparov knew all the best moves, so it was just a matter of programming them into the machine. Decades later, Google created another chess playing machine, but this computer learned by observing. It didn't use Boolean logic like my plastic box. Instead, it watched experts play games. Then it used probability to make guesses and learn from its mistakes. This technique was called machine learning. This newer type of artificial intelligence could do much more than just play chess. It could beat humans at complex video games. These games required real-time creativity and movement. Google used this machine in a game called, "Go", a game so complex that it could never be programmed using simple IF-THEN statements. There's simply too many moves and counter moves. The Go machine used something called a deep learning artificial neural network. This was a form of machine learning that allowed it to process enormous amounts of data about strategy. This deep learning machine was able to come up with strategies that humans never imagined. In fact, when it played against the world champion ago, the machines moves were so unusual that most people thought it was losing, until it won. I half imagined my old gravely flight companion leaning over the Go master and saying, "I guess that thing's smarter than you are." In this case, he'd be closer to right. Newer machine learning systems move beyond simple logic. These deep learning systems start to blend into human domains, such as creative problem solving. These newer deep learning systems could use probability to make accurate predictions. It could also make guesses about the future. When its opponent made a move, it would shift its probabilities and make a new set of predictions. Even though it started with games, predictive deep learning systems are now all about business. These systems created tremendous value as they got better at making extremely accurate predictions.