Slashdot Mirror
Facebook AI Director Discusses Deep Learning, Hype, and the Singularity
An anonymous reader writes In a wide-ranging interview with IEEE Spectrum, Yann LeCun talks about his work at the Facebook AI Research group and the applications and limitations of deep learning and other AI techniques. He also talks about hype, 'cargo cult science', and what he dislikes about the Singularity movement. The discussion also includes brain-inspired processors, supervised vs. unsupervised learning, humanism, morality, and strange airplanes.
There is a point where the first marginal barely even an AI, wouldn't win any Turning contest, largely useless AI will be created. But if the algorithm is evolutionary in nature it could be the point where it then improves itself, then improves itself, and so on until pretty much out of nowhere you have an indisputable AI.
I regularly employ genetic algorithms and can say without hesitation that I have little idea how they got to where they got and the results are often fantastic. But my code is usually a single layer. That is I have a target, I set up the parameters it needs to explore, and then I set it loose. This is because the number of permutations exceed what my computer can handle in a reasonable time (a-la travelling salesmen problem) and a GA will get me close enough much faster.
But if I added a second layer where the GA was noodling with my code then I suspect interesting things could happen; not an AI but I doubt that I could comprehend the code it would generate. This will be the route to an AI. Basically the key will be an algorithm that generates not only code that we can't comprehend but generates the next generation of the GA which generates another generation of the GA and so on until we have code so far removed that when it works it will be just like where we are with understanding the overall design of the brain (we largely don't).
What this boils down to is that I very much doubt that AI will be the step by step process like most of human endeavour where we can see it coming but one where it is like trying to open pandora's box "just a crack". One day we will have an interesting algorithm and that evening we will have AI. Sort of a directed emergent property.
One other bet is that it won't be an "AI" researcher who will build it. It will be someone working on some other NP hard algorithm such as protein folding or image recognition.
To me the only question is one of math. Is there a minimum processing power required for an AI that can deal with a real time universe? At that point we can at least calculate when we might have an AI that is something that needs to be dealt with. I am also fairly certain that the moment we cross that computational threshold an AI will soon follow.