Clickbait.
ML won’t require mass amounts of data forever. One-shot learning will get solved.
ML won’t require mass amounts of data forever. One-shot learning will get solved.
This is nonsense. I would challenge you to show any few-shot learning work that isn't basically some form of transfer learning in disguise.
Few-shot learning is super useful if you're in a position where you don't have much data, but will never compete with huge stacks of data.
>will never compete with huge stacks of data
Children learn with sparse data based on composable models, and are bio machines.
"I would challenge you to show proof of X" in ML is an empty challenge when the technology is in its infancy. You might as well say "I challenge you to stream HD video over the internet" in 1989. Theoretically possible but technically impossible for the time. We proved it possible as tech improved, though.
The problem with cynicism is it costs nothing but still makes you look like an expert.
This has literally nothing to do with few-shot learning though. You can always make a crappy model with a small amount of data and then improve it as you get more. And my point is that if you have a competitor with several orders of magnitude more data, their models are almost certainly still going to be better than yours.
Few-shot learning will probably be able to improve the baseline of what you can do on certain tasks, but you're not magically going to learn a sufficiently accurate cancer diagnosis algorithm from 5 radiology images.
The best case for few-shot learning is that you go from "worse than the alternatives" to "good enough for some people to get value", which will probably happen a few times, but is going to be a minor phenomenon.