3 ms·

Clickbait. ML won’t require mass amounts of data forever. One-shot learning will get solved.

Eridrus 9y ago
> 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.
- bobsil1 9y ago
  Of course it's a kind of transfer learning. >will never compete with huge stacks of data Children learn with sparse data based on composable models, and are bio machines.
  - andreasgonewild 9y ago
    This boils down to a fundamental divide in life view. Anyone who has done any amount of expanding their consciousness will tell you that there is a ghost in the machine. It's always a mistake to assume that all you can see is all there is to see.
    - bobsil1 9y ago
      I am a materialist, there is no magic :)
      - Eridrus 9y ago
        I agree that there is no magic, but evolution has done some pretty impressive initialization.
      - andreasgonewild 9y ago
        You are also most probably human, have you even bothered to have a look inside?
    - asadlionpk 9y ago
      I do some AI work and recently have been thinking on this, can you expand on this. Or point me to the right direction regarding this idea.
      - andreasgonewild 9y ago
        It's always there for anyone to grab. Try lying on your back with all your gadgets turned off and eyes closed; and observe your brain thinking for a while; which is about as difficult as meditation needs to be. Might take some practice to not identify with the thoughts and get carried away, or to not block yourself by trying too hard. Then ask yourself while lying there: Who am I? Literally :) Who is observing the thoughts? The experience is not of this world, it can't be transferred using words. There are other ways to get a glimpse, mushrooms, ayahuasca etc; but introspection is the way to go for lasting results. I spent 32 years writing software, and I just know that a computer will never be able to do the same thing.
- freehunter 9y ago
  The nice thing with digital models is they can get data from everywhere and continue to improve slowly over time. It's the lean startup applied to ML: there's nothing wrong with launching with a one-shot ML product to test a hypothesis, and as time goes on keep adding to it. That's the core concept there: start with a working but basic model and continue to extrapolate from that as time goes on. "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.
  - Eridrus 9y ago
    > That's the core concept there: start with a working but basic model and continue to extrapolate from that as time goes on. 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.
justonepost 9y ago
Yeah, AlphaGo Zero is the tip of the iceberg.

> 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.

bobsil1 9y ago
Of course it's a kind of transfer learning. >will never compete with huge stacks of data Children learn with sparse data based on composable models, and are bio machines.
- andreasgonewild 9y ago
  This boils down to a fundamental divide in life view. Anyone who has done any amount of expanding their consciousness will tell you that there is a ghost in the machine. It's always a mistake to assume that all you can see is all there is to see.
  - bobsil1 9y ago
    I am a materialist, there is no magic :)
    - Eridrus 9y ago
      I agree that there is no magic, but evolution has done some pretty impressive initialization.
    - andreasgonewild 9y ago
      You are also most probably human, have you even bothered to have a look inside?
  - asadlionpk 9y ago
    I do some AI work and recently have been thinking on this, can you expand on this. Or point me to the right direction regarding this idea.
    - andreasgonewild 9y ago
      It's always there for anyone to grab. Try lying on your back with all your gadgets turned off and eyes closed; and observe your brain thinking for a while; which is about as difficult as meditation needs to be. Might take some practice to not identify with the thoughts and get carried away, or to not block yourself by trying too hard. Then ask yourself while lying there: Who am I? Literally :) Who is observing the thoughts? The experience is not of this world, it can't be transferred using words. There are other ways to get a glimpse, mushrooms, ayahuasca etc; but introspection is the way to go for lasting results. I spent 32 years writing software, and I just know that a computer will never be able to do the same thing.
freehunter 9y ago
The nice thing with digital models is they can get data from everywhere and continue to improve slowly over time. It's the lean startup applied to ML: there's nothing wrong with launching with a one-shot ML product to test a hypothesis, and as time goes on keep adding to it. That's the core concept there: start with a working but basic model and continue to extrapolate from that as time goes on. "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.
- Eridrus 9y ago
  > That's the core concept there: start with a working but basic model and continue to extrapolate from that as time goes on. 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.

justonepost 9y ago

Yeah, AlphaGo Zero is the tip of the iceberg.