Anaxagoras wrote: ↑
Wed Dec 12, 2018 3:47 pm
Here's a thought. This seems like something that an AI could do in principle, and it doesn't even need to have free will or be able to think "like a human" necessarily to do it:
Can it do a decent translation from one language to another? As good as a good human translator. There are of course plenty of terrible translators out there, but I would like to know when AlphaZero can translate say a novel from Japanese to English or vice versa. It would take a lot of work, but it seems like the kind of thing that an AI like AlphaZero could do, if it had the right kind of data to learn from. Can it teach itself to do it? I sort of doubt it, but maybe a clever person could figure that out. Last time I checked, Google Translate still kinda sucked for Japanese to English. I check every once in a while. I haven't tried other languages. Maybe someone can try German to English or French to English or Russian to English and see how well it does.
That's a great thought.
I don't know enough about AlphaZero to comment about its abilities, but I can point out that in general, there is a bigly and fundamental difference between AI learning to play chess versus translating languages. And that is the feedback to know how well you did. With chess, the AI can play against itself (or anything else) and evaluate the results without any human intervention. It is in essence self correcting with enough experience. It cannot do that with translating languages since there is no similar feedback mechanism without human intervention. Or to put it another way, with chess, it can create its own knowledge how best to play the game. It cannot do that with translating languages, since it needs humans to tell it how well it did.
On a related note, I have a hypothesis about why AI does not (currently) seem to do well in classifying music genres. It depends on examples of human classifications to learn from. My conjecture is that the AI classifications are probably very good, but that the human classifications are faulty, and that might explain why AI doesn't seem to do well. For example, the AI is given a largish database of music which is pre-labeled by genre according to humans. The AI then tries to figure out the rules of classification from that database. My hypothesis is that humans are not very good at that, and thus the AI is hamstrung before it even gets started.
Getting back to the chess example, the AI can create its own database of winning and losing moves. With translating languages, it cannot create its own database of good and bad translations. At least not with the current state of the art, as I understand it.
Or to point out one more obvious difference, evaluating the results of a chess game is largely objective, compared to language translation which has a large subjective component.
I could be wrong. But it's an interesting question.