You can handle arbitrarily complex tasks with large parametric models trained with SGD. The problem is that doing it well requires a *dense sampling* of the input/output space you're learning, because the generalization power of these models is extremely weak. That's expensive.
Deep learning is immensely useful, but it does not bring meaningfully closer to understanding intelligence (as for "AGI", that's a sci-fi talking point, as even human intelligence is specialized). Intelligence is 100% about efficient generalization. DL is orthogonal to that.
Schematically, intelligence is skill divided by experience (I = S/E). Deep learning enables arbitrarily high skill levels, but requires insanely high amounts of "experience" (data) to achieve these levels, resulting in an extremely low intelligence factor.
Again, a DL model requires a dense sampling of what it's doing. An intelligent agent (like a human) can do extreme generalization from little data. At this time, no one has any clue how that works. However, it may not necessarily be very complicated. Who knows...
You can follow @fchollet.
Tip: mention @threader_app on a Twitter thread with the keyword “compile” to get a link to it.
Enjoy Threader? Sign up.
Threader is an independent project created by only two developers. The site gets 500,000+ visits a month and our iOS Twitter client was featured as an App of the Day by Apple. Running this space is expensive and time consuming. If you find Threader useful, please consider supporting us to make it a sustainable project.