Mike in Analysis 3 minutes Providence

AlphaGo and Deep Learning

AlphaGo is making news for a five-game match against top go player Lee Sedol.

AlphaGo is a go computer engine built with deep learning. A vast number of games are fed into a neural network, and we let the program learn. Is this the promised future? Don’t solve a problem, instead: define a successful outcome, give some initial data, and let the computer solve it. My friend Fran put it this way: “because AI programming it’s going to be more about building the teaching material than anything else, we’ll be able to replace programmers with just teachers.”

We’d be more excited about AI advancements if we had a more active role. In computer science we rely the idea of tightly controlling the input, the output and the logic in between. With deep learning, we’re starting to cross Clark’s technology/magic barrier.

Put another way, the study of AI promises the chance to understand our own intelligence. Deep learning robs us of any such potential insights, as the computer is doing all the hard work.

On the other hand, it’s very like us to employ technology before we understand it. Military and private industry will likely refine the process of creating deep-learning black boxes to the point that they are very sharp and reliable tools.

I predict that the old study of AI will be relegated to some dusty academic research wing of a brick and mortar university, who might even start to attempt to “reverse engineer” some black-box AI, developing a new field of “Computer Program Psychology”. <grin>

Methodology aside, it’s amazing they’ve managed this, and I am looking forward to seeing the outcome of AlphaGo’s battles with Lee Sedol!