We’re moving more and more into the age of machine learning – whether it’s semantic engines that extract entities from free text or algorithms that attempt to divine market sentiment from a flood of tweets or programs that can turn out stories from SEC filings. Sometimes it seems like magic. Or the end of the world. Or both.
The New York Times has built a great app that shows, in a very visible way, just how machines can learn. It’s a rock-paper-scissors game that you play against the computer – and it learns from you as well as from a store of all the other games it’s ever played. It’s addictive and educational – a nice combination.
Play a couple of rounds at the novice level, then click on the button that shows what the computer is thinking. The first thing you learn is that the most basic rules haven’t been programmed into the app – it only “learns” when it wins, for example, that rock beats scissors. Then after the while it starts to discern patterns in your play – even if you think you’re playing randomly – and can try to play against it.
But it’s at the veteran level that the program shows its power. Or more specifically, the power of statistics and big numbers. You don’t have to program the computer with rules on how to win at rock-paper-scissors; you only need to let it iterate through enough games so it learns from strategies that have worked in the past – and at the veteran level, it brings in all the previous games it’s ever played: thousands of them.
Years ago – three decades ago, when I was in university – I programmed a simple tic-tac-toe game on much the same lines: No rules other than an ability to learn from winning strategies. Unfortunately I didn’t have time to play enough games against the computer (or more precisely, I set it to play randomly against itself) – which is the result of waiting until the last minute to finish your homework assignment. So it didn’t play brilliantly, but it more of less worked as an example of an artificial intelligence learning algorithm. And more importantly, I passed the class.
That program was far less sophisticated than this one, but many of the principles are the same – one key difference these days is that we have access to much more data, as well as the opportunity, via the internet, to run programs like this on much bigger scales. Thanks to the age of big data, machines have the opportunity to learn that much more quickly and that much more precisely.
We need to engaged in this process – partly because it can provide us with potential troves of information and insight; but partly also because others are using these techniques to extract value out of what we do, so we should at least understand how it’s happen. Or we’ll be left behind by another wave of technological change.