4 ms·

I don’t believe this statement is true at all. One can learn simply by writing and receiving feedback on their writing. The training corpus is merely a helpful

I don’t believe this statement is true at all. One can learn simply by writing and receiving feedback on their writing. The training corpus is merely a helpful starting point. You could easily build a good model on model outputs. Building a model on curated outputs may well be more useful than training a model on a collection of human outputs tainted by human stupidity for some purposes. Already we have a decent ability to coverage new functionality these models by chatting rather than retraining. So long as a generalizing capability is built into it.

1shooner 4y ago
> Building a model on curated outputs may well be more useful than training a model on a collection of human outputs How is the curation not a human output?
- brookst 4y ago
  If it’s curated by AI?
- meh8881 4y ago
  It could be. The difference between needing a human to rate a thing vs create a thing is very different though
tpoacher 4y ago
This is less about 'self-supervised' learning, and more about ground truth. I see a variant of this in medical AI. E.g. people generate 'fake' augmented Brain MRI datasets using image processing tricks, in order to provide more training data. But the 'fake' MRI datasets are clearly 'similar looking images', which may or may not be anatomically correct. If you start accepting a flood of fake images as legitimate data for training AI, then it will be impossible to trust the predictions, however good they are in the majority of cases.
- brookst 4y ago
  I think you’re describing introduction of noise to training sets, which is a staple of training. Your definitions of “fake” and “legitimate” are circular, and miss the central point of large ML models: they can extrapolate from imperfect data because of the massive scale. Yes, the predictions will be imperfect. That’s true today, of both ML models and human radiologists. It’s about reducing the error rate, not designing a perfect algorithm that is never wrong. I’m pretty sure Gödel or someone can explain why the later isn’t even possible, for machine or human.
- meh8881 4y ago
  That does not feel like a good example. The context here is largely generative models relaxing human creations. We don’t need to generate Brain MRIs. This use case you outline is a niche thing for trying to train better models. Not doing the thing we actually need to do at such a scale that humans aren’t doing the original task anymore.
brookst 4y ago
Yep. We’re already training models based on ML models of what human feedback would be for a response. I see no difference between that and training models on inputs that a model estimates are desirable to humans. This is classic “processor design will hit a wall when they get too complex for humans to understand every gate” criticism: it assumes limits on a tech advance, when the limits themselves are being wiped out by the advance.
- still_grokking 4y ago
  Sole humans can still trace every gate on a modern chip, or even the gate's components, and how exactly they work, to some explicit specification. It's completely deterministic and a human can fully understand all details. Otherwise it wouldn't be possible to debug it, or manufacture it in the first place. OTOH we can't track or understand the results of AI. It spits out results but we can't understand how exactly it deduced those results, usually. That's the unfortunate current state actually. So the analogy doesn't hold. But this seems not even relevant. Because no level of technological progress will invalidated the old "garbage in, garbage out" principle. If you make a loop out of it, there's more or less only one thing to expect… An AI would need to know how to teach itself "reasonable things". But to be able to do so it would need human level intelligence. Growing AI is likely like growing children: Children can't teach themself in the beginning. You need to teach them some basics before they (maybe) learn how to teach themself. AI is not even near to "know" any basics. So it will need a lot of well thought out human input before it can proceed to teach itself.
  - meh8881 4y ago
    Writing garbage, then receiving feedback on your writing, and progressing to write less garbage is a common pattern of improvement however. The origin of the corpus matters much less than the quality of its curation.
speleding 4y ago
I agree, if the model is powerful enough one should be able to ask "Please find the errors in this text book", and then feed the corrected version back into the training data. Rinse and repeat for lots of books.
ismokedoinks 4y ago
Wouldn't that just create simulacra of human language? Seems difficult to draw a circle around "human stupidity" while keeping the innate human qualities & also not amplifying similar AI aberrations

1shooner 4y ago

> Building a model on curated outputs may well be more useful than training a model on a collection of human outputs How is the curation not a human output?

brookst 4y ago
If it’s curated by AI?
meh8881 4y ago
It could be. The difference between needing a human to rate a thing vs create a thing is very different though

tpoacher 4y ago

This is less about 'self-supervised' learning, and more about ground truth. I see a variant of this in medical AI. E.g. people generate 'fake' augmented Brain MRI datasets using image processing tricks, in order to provide more training data. But the 'fake' MRI datasets are clearly 'similar looking images', which may or may not be anatomically correct. If you start accepting a flood of fake images as legitimate data for training AI, then it will be impossible to trust the predictions, however good they are in the majority of cases.

brookst 4y ago
I think you’re describing introduction of noise to training sets, which is a staple of training. Your definitions of “fake” and “legitimate” are circular, and miss the central point of large ML models: they can extrapolate from imperfect data because of the massive scale. Yes, the predictions will be imperfect. That’s true today, of both ML models and human radiologists. It’s about reducing the error rate, not designing a perfect algorithm that is never wrong. I’m pretty sure Gödel or someone can explain why the later isn’t even possible, for machine or human.
meh8881 4y ago
That does not feel like a good example. The context here is largely generative models relaxing human creations. We don’t need to generate Brain MRIs. This use case you outline is a niche thing for trying to train better models. Not doing the thing we actually need to do at such a scale that humans aren’t doing the original task anymore.

brookst 4y ago

Yep. We’re already training models based on ML models of what human feedback would be for a response. I see no difference between that and training models on inputs that a model estimates are desirable to humans. This is classic “processor design will hit a wall when they get too complex for humans to understand every gate” criticism: it assumes limits on a tech advance, when the limits themselves are being wiped out by the advance.

still_grokking 4y ago
Sole humans can still trace every gate on a modern chip, or even the gate's components, and how exactly they work, to some explicit specification. It's completely deterministic and a human can fully understand all details. Otherwise it wouldn't be possible to debug it, or manufacture it in the first place. OTOH we can't track or understand the results of AI. It spits out results but we can't understand how exactly it deduced those results, usually. That's the unfortunate current state actually. So the analogy doesn't hold. But this seems not even relevant. Because no level of technological progress will invalidated the old "garbage in, garbage out" principle. If you make a loop out of it, there's more or less only one thing to expect… An AI would need to know how to teach itself "reasonable things". But to be able to do so it would need human level intelligence. Growing AI is likely like growing children: Children can't teach themself in the beginning. You need to teach them some basics before they (maybe) learn how to teach themself. AI is not even near to "know" any basics. So it will need a lot of well thought out human input before it can proceed to teach itself.
- meh8881 4y ago
  Writing garbage, then receiving feedback on your writing, and progressing to write less garbage is a common pattern of improvement however. The origin of the corpus matters much less than the quality of its curation.

speleding 4y ago

I agree, if the model is powerful enough one should be able to ask "Please find the errors in this text book", and then feed the corrected version back into the training data. Rinse and repeat for lots of books.

ismokedoinks 4y ago

Wouldn't that just create simulacra of human language? Seems difficult to draw a circle around "human stupidity" while keeping the innate human qualities & also not amplifying similar AI aberrations