Thursday, March 17, 2016

Machines that Think

Michael - Thursday

It’s a favorite theme in Science Fiction that thinking machines lie in our future, some for good, some for ill. Remember Hal 9000 in 2001 - A Space Odyssey? The moral of the story always seemed much deeper than the technology. It seems to be that creating a thinking being – no matter of what material – is God’s provenance and we trespass on that at our peril. Well, we’re finally catching up with the SF. No less a person than Stephen Hawking has warned of the dangers to humanity of intelligent machines.

And yet so called Artificial Intelligence - once the hottest area of computer science - seemed to over claim and under deliver. There are a number of benchmarks out there and they’ve been hard to reach. One was Turing’s test. The idea is that you talk over the phone to either a computer or a human and you cannot distinguish between them.  That challenge stood unfulfilled from Turing’s time until a few years ago. And yet this is not the test of a thinking machine. It’s a test to see if a computer can be programmed to impersonate a human. That’s very different.  (Not that it’s not hard. Smart as she is, chat to Siri on your iPhone for a few seconds. No one would be fooled into thinking it’s a real woman.)  Another benchmark was chess. Chess is a difficult game with a well-defined set of rules and chance isn’t involved. The skill is in thinking ahead and recognizing that one position will be better than another.  The target was to design a computer which could beat a chess grandmaster.  

IBM’s Deep Blue did so twenty years ago beating Gary Kasparov in the second of their match series.. It was a super achievement. The trick was to evaluate the various positions in terms of a favorability analysis, and then provide the computing power and memory which allowed the machine to think further ahead than its opponent. Impressive, BUT let’s analyze this “thinking ahead.” It’s not thought at all. It simply involves flowing the possible sequence of reasonable moves from where you are now into the future. This isn’t trivial. There may be around 40 moves available to each side in any position. Some will be rejected almost at once, so say there are 10 realistic moves. After you and your opponent have made five moves each, you have reached over 10 billion possibilities! Human grandmasters often think ten moves into the future of the game, but they don’t evaluate 10 billion different positions. They see that only a few of those are worth following because the others will clearly not be to their advantage.  Clear to them… 

Over the last few years there’s been a renaissance in artificial intelligence under the banner of “deep learning.” With humans, deep learning means the type of learning achieved by working through the issues and seeing where they lead. Trying techniques, making mistakes, learning from the mistakes, internalizing the knowledge and ability. It’s a very different sort of approach to learning by rote.

Think about Deep Blue’s remarkable achievement. At the crudest level it was a matter of computing many possible situations and evaluating them by means of matching to good and bad templates produced for it by humans. This is not learning at all.

In the machine learning world, deep learning means the ability of the computer to do the same sort of thing we talk about in human learning. Given the rules of the game it can arrange its own “neurons” to reflect the rewards it gets for good moves and the penalties it gets for bad moves. This mimics deep learning in people – at least at some level.

Last week deep learning had its greatest triumph so far. Google has a whole division – Deep Mind - working on it. They designed a computer to play Go.

Go is regarded as a much more complex game than chess. It offers the players hundreds of different moves in each position and the efficacy of a move is unclear for quite a way into the future. Deep Blue doesn’t know how to play Go, and if it did it would lose. We don’t understand the game well enough to do good evaluations of the positions and we can’t play the moves far enough into the future. Ten moves into the future would lead to 100 billion billion possibilities! Forget about it.


The Google machine – AlphaGo – already had a pretty impressive record of success. But over the last week it beat the world champion Lee Sedol by 4 games to 1 in a five match $1 million series. This is a thinking machine. It taught itself to play Go. The Google programmers don’t know how it does it.
After AlphaGo won its first game, the New York Times carried an article made up of four short pieces from experts in different aspects of computer science. You can read the whole piece HERE. Interesting that much of the emphasis is on how humans should respond to this new development. 

Here are a couple of quotes:

Google’s AlphaGo is demonstrating for the first time that machines can truly learn and think in a human way…The real question, now that machines are capable of approximating human intuition in decision making, is: How should we cultivate human talents going forward? Because it's clear that the human advantage is eroding fast - Howard Yu, professor of strategic management and innovation.

The third critical [human] ability, somewhat surprisingly, is storytelling, which has not traditionally been valued by organizations. Charts, graphs and data analysis will continue to be important, but that’s exactly what technology does so well. To change people’s minds or inspire them to act, tell them a story - Geoff Colvin, a senior editor at large for Fortune Magazine.

Of course I couldn’t resist that second one! Maybe writers will be the new elite! I won’t hold my breath.

Another of the experts points out that it’s a long way from learning to play a formal game with fixed rules and no downside for mistakes to thinking in the uncertain real world environment in which we operate every day. And that brings us back to HAL. It wasn’t thinking that was the issue there, it was consciousness. And that’s a whole other ball park.

9 comments:

  1. Thanks, Michael. Of course, there's a whole WORLD of debate about what, exactly, is 'consciousness.' But ignoring that, we (humans) seem to finally be approaching the ability to design systems that can learn AND design. It works best on a different kind of hardware than 'modern' computers (which are VERY 'precedural' oriented). But, as we dance around the problem from many sides and keep getting closer and closer, the issue that is concerning many people is how do you control it? Once you reach near "human level" artificial intelligence, the 'machine' can then design new versions of itself, bootstrapping its own intelligence to higher and higher levels. There's a whole area of study on how to keep such an artificial intelligence from 'escaping' and running away from our control, which is what scares folks like Stephen Hawking and Elon Musk. The benefits of such an artificial intelligence are immense. The risks are equally immense (for humans :-).

    But I'm an optimist. In the game of life and death, the pessimist is wrong every day, the optimist is only wrong once...

    ReplyDelete
    Replies
    1. I love that last quote! Yes, I'm also an optimist. I think there are dangers, but nothing compared to global warming for example. (Fill in your favorite man made catastrophe!)

      Delete
  2. Michael, this is so thought provoking. And, as ever, a bit scry. Should I be comforted but the fact that computers need electricity to operate. They die if that is interrupted for a second. Humans can get along quite well without it. Of course, if it went away for good, a lot of us would die, so I hope it doesn't. But we could always cut them off from their "air supply." Right?

    ReplyDelete
    Replies
    1. This was one cartoonist's suggestion for the Kasparov/Deep Blue match!
      It shows Kasporov's hand reaching for the power switch and the caption is 'Kasparov beats Deep Blue in one move'!
      However, robots tend to be built with long lasting batteries...

      Delete
  3. Can a machine ever develop the flaws that make us human and imperfect in a random way?
    The other ball park is Wrigley Field. I'm pretty sure.

    ReplyDelete
  4. For the man made catastrophe I'd chip in the amount of hormone in the food change and then antibiotic resistant infection. The infection will be caught from a computer keyboard!

    ReplyDelete
  5. From the level of Presidential US candidates I'd say much of the country already has lost consciousness and is primed for a robotic takeover.

    ReplyDelete