So I finally got around to watching Caprica, the spin-off pilot from the mostly-wonderful Battlestar Galactica. As a scientist, I'm usually pretty happy to just let the science-fiction bits of such programs wash over me: to just suspend disbelief and treat them as fantasy. With Battlestar this was particularly easy to do, as so little of the science was really explained. But with Caprica, the way they said you could create a virtual human personality was so plausible, well-explained, and—in my view—so completely wrong, that I thought it worth putting something on the record.
So the idea is this. If you take all the data about us that potentially exists—from photos to medical records, to Google searches, to school results (supposedly about 300Mb)—you can then infer everything you need to know in order to recreate our personalities. This is an absolutely terrific plot device if you're a TV writer, because it means that you can recreate a person without having their permission or ever having them present. Also, it sounds plausible because, if you've been reading up on deep data mining, you'll know that information from one domain can indeed be used to make inferences in other domains. (Where you bank may imply you're more likely to wear brown shoes).
However, I would argue that—if we've learned anything at all about how our brains work over the last 30 years or so—it's that, as machines, we're driven by association, not data. Synapses are strengthened when neurons fire together and wither when they don't. We learn because experience has taught us associate sensory inputs—what we see, hear taste, touch, and feel—with positive or negative outcomes. These may be immediate, or steps along a chain of events. And the association may be with the exact stimulus or something that's just similar. So I might be especially nice to someone who looks a bit like a favorite teacher I once had and not even realize why I instantly liked her. It's associations like this—to people, to food, to situations—that encode personality. Knowing the grades I got in my teacher's class would help you not one bit in predicting how I would respond to the new face.
This may sound reminiscent of the debate between 'good old-fashioned AI', which is basically top-down and knowledge based, and behavioral or associative AI. Well, it's not really a debate any more: books like Jeff Hawkins' On Intelligence succinctly explain that animal (including human) intelligence is much more about pattern matching than anything else. Which is not to say that it may not be expedient to use expert-system-type systems within an AI you are building: its just that such systems will not get you very far if you want to build a human-like AI.
Oh well. No doubt I'll soon be as addicted to Caprica as I was to Battlestar: apart from my grumbles above, I found it completely compelling. Let's just hope that the amount of tech-speak in it drops off enough that I don't have to turn off my entire left brain to get through it... or embarrass myself by shouting at the TV.
I tend to moan about surreal travel times and artificial gravity but I've also learn to put it aside and enjoy the show.
I think they're better off being a complete non-sense than trying to explain things. It's easier to shut down the brain and kinda makes you feel deeper into the story...
Posted by: Renato Golin | Sunday, 10 May 2009 at 10:08 PM