Geoff Hinton on AlphaGo and the Future of AI
macleans.ca
macleans.ca
What neural network have demonstrated so far is that it can do a form of perception given enough ground truth data. Speech recognition and image recognition are all form of perception. Go playing is also mostly about perception: recognizing good patterns.
There are multiple issues with extrapolating this achievement to anything more than what it is.
First, perception is a only minuscule part of the brain's functionality. One only needs to glance at the table of content page of any cognitive psychology 101 textbook to see that. Second, it is not clear human's perceptual ability is the same as what NN has demonstrated. For example, human's perceptual learning is mostly based on unsupervised learning, not supervised like in NN.
Finally, his dismissal of "traditional" approach is extremely short-sighted. He seems to suggest that we can solve AI by just throwing more hardware and power at NN. This view is very narrow minded.
He also seems to overlook the high possibility that the "traditional" approach could benefit similarly from the advances in hardware and power. I predict the next breakthrough in AI will come from that direction.
His comments on his own professional experience, toward the end, are also very interesting -- not the sort of discussion one usually sees from senior scientists.
[on rendering humanity obsolete]
A: It’s hard to predict beyond five years. I’m pretty confident it won’t happen in the next five years, and I’m fairly confident that it won’t be something I’m going to have to deal with. But it’s something people should definitely be thinking about.
OK kids, were making AI -- deal with it!!!
FWIW I agree with him on the time frame -- we're in another phase where AI successes have led to overoptimism. The human brain will humble us in the next decade....
I get the sense that some machine learning/AI researchers downplay this long-term concern, while at the same time misattributing discussons around AI safety as criticism of the current approach to AI.
It must be possible to have two thoughts in our heads at once. Machine learning is making massive strides at the moment, and we must use this for all the benefit it's worth while keeping an eye on the current risks that AI represents -- for instance job loss and displacement, misuse by governments for e.g. surveillance purposes and so on. But it's probably also a good idea to keep one eye on the horizon as well, because things are likely to happen pretty quickly if the state of the art advances further.
It's best to be prepared for this eventuality well in advance, because successful navigation of a potential risk scenario there requires the cooperation of a lot of people. And people are slow to change their expectations.
Don't get me wrong, I greatly appreciate the current discussions around this subject. There are a lot of good points being made.
From basic game theory, each board game such a Go, chess, checkers, Nim, has an optimal strategy. Then the only issue left is who moves first. If the advantage is to the player who moves first, then they should always win; similarly for the player who moves second. Details in, say,
T. Parthasarathy and T. E. S. Raghavan, Some Topics in Two-Person Games.
Well, such a strategy is known for Nim, e.g.,
Courant and Robbins, What Is Mathematics?
So, when the strategy is known, then following the strategy stands to beat Hinton's AI. Or, Hinton's million node AI can learn and learn all it wants, millions of games, maybe against itself, and just a tiny computer program, if it has choice of who moves first, will always win (no ties in Nim).
More generally, for now, when can solve a cleanly specified problem mathematically, that is, complete with theorems and proofs, then that is better than Hinton's AI.
Maybe in time something like Hinton's AI will be able to do research in math and produce results as theorems and proofs, maybe find the optimal chess strategy, and then, with choice of who moves first, never lose at chess, etc.
But here's
Lesson 1: For now, when math can solve a problem, it is better than Hinton's AI.
So, sure, for startups, here's
Lesson 2: F'get about Hinton's AI. Instead, pick a cleanly specified problem where can do some math research and get a solid math solution. Then stand to beat intuition, Hinton's AI, and, in practice, about everything else.
Are there some problems now where math can do better than Hinton's AI? Yup.
Point 1: Whatever can be done with Hinton's AI broadly in the economy, one of the keys to a successful startup is being exceptionally good. Well, a way to never lose and usually to win, usually win big, is to have the best math solution. So, that's one of the criteria in picking what a startup is to do. Again, it's just crucial to be exceptional.
Simple example: Hinton's AI would have one heck of a time competing with the simplex algorithm on realistic problems, say, from running an oil refinery, on linear programming. A worse time competing on non-linear programming, say, also for running an oil refinery. A worse time competing on stochastic optimal control. Really hopeless on competing with Issac's differential game theory in pursuit-evasion games. And there's lots more.