Unsupervised learning: the missing link of AGI(technologyreview.com) |
Unsupervised learning: the missing link of AGI(technologyreview.com) |
I don't know about that. I think it's safe to say that the way humans learn is nothing like the way computers learn, to the point that such comparisons are completely meaningless. We don't get given covariance matrices to crunch through, in order to learn how to walk or talk, say. There's a process, sure, but it seems to be very complicated and certainly nothing like the very simple training we use in machine learning.
You can see examples of this all over the place. Frex, recognising cats: by the time a human child is able to tell a cat from a dog, it's also able to recognise a bazillion other things, including language- even multiple languages. Whereas a machine learning algorithm that learns to recognise cats can only ever recognise cats- if you want it to recognise dogs, you have to train it from scratch, on an entirely different dataset. And even if you don't have labels, you still have to curate the dataset, by hand. Someone has to tell the bloody thing what a cat looks like, even if nobody tells it that it's called a "cat".
And there's the rub- humans learn a lot from very little data, with very little help. Machine learning has to slog through untold bytes of data, painstakingly manipulated by a human, just to learn a measly, puny little thing that's only ever useful in a very limited context.
So basically there's no comparison. Supervised or not is not the point, the point is that our current algorithms are extremely limited in what they can learn and we're not going to get AGI out of them, one way or another.