|
|
|
|
|
|
by PaulHoule
97 days ago
|
|
|
... well, well, well. I spent a lot of the 2010s revisiting symbolic AI and I'd say the worst problem it had was "reasoning with uncertainty" If you consider a medical diagnosis system like https://en.wikipedia.org/wiki/Mycin the result is probabilistic in nature, there's always some chance you'll get it wrong. Language processing is the same. Language is ambiguous, there are thousands of possible parse trees for a common sentence. You might be talking with somebody and then get a piece of information that revises your interpretation of what they said an hour ago. It's just like that. In that time frame I was very interested in the idea that decision theory was the key link between computation and action whether you were using symbolic methods (e.g. a very plausible set of rules for address matching might be 99.9% reliable in some cases, 97% in others, 2% in others) or learned methods. A model for predicting market prices is priceless, but put that together with a Kelly Better and you've got a trading strategy. Maybe there is more to his argument than I got, but as I see it he's defending a boundary that isn't there. |
|
|