[HN Gopher] Will we run out of ML data? Evidence from projecting...
___________________________________________________________________
Will we run out of ML data? Evidence from projecting dataset size
trends (2022)
Author : kurhan
Score : 29 points
Date : 2023-04-22 15:50 UTC (7 hours ago)
(HTM) web link (epochai.org)
(TXT) w3m dump (epochai.org)
| cs702 wrote:
| No.
|
| AI can generate as much synthetic data as we need, on demand.
|
| Many SOTA models, in fact, are already being trained with
| synthetic AI-generated data.
|
| See https://en.wikipedia.org/wiki/Betteridge's_law_of_headlines
| Buttons840 wrote:
| > AI can generate as much synthetic data as we need, on demand.
|
| I don't think this is right.
|
| Can I take an untrained LLM (a neural network with random
| parameters), and have it start generating garbage, and then
| train the network to produce more of the same and then have it
| bootstrap itself to intelligence? Of course not.
|
| What if I train it just a little bit first? What if I train it
| until it produces gibberish, but does occasionally string two
| words together that are spelled correctly. Can I have it
| produce petabytes of gibberish and then train on that to reach
| GTP4's level?
|
| You seem to argue that at some point, the AI is able to improve
| by training on its own output. At what point does that arrive?
| Because so far we've never seen an AI improve based on its own
| output. (As far as I know?)
| blackbear_ wrote:
| > Because so far we've never seen an AI improve based on its
| own output.
|
| Maybe it's because AI is such an overloaded term, but this is
| pretty commonplace for (semi-)supervised learning algorithms.
|
| Pseudo-labeling [1,2] is an example of this that has been
| around for decades. When done properly it does improve the
| performance of the original model, up to a certain limit (far
| from the singularity).
|
| Moreover, it is apparently possible to improve a model's
| performance by augmenting it's training set with synthetic
| examples generated by a second model [3].
|
| Finally, boosting [4] can also be seen as iteratively
| leveraging the output of a model to train a slightly better
| model. In fact, a specific type of boosting often yields
| state of the art performance on tabular data.
|
| [1] https://arxiv.org/abs/2101.06329
|
| [2] https://stats.stackexchange.com/questions/364584/why-
| does-us...
|
| [3] https://arxiv.org/abs/2304.08466
|
| [4]
| https://en.m.wikipedia.org/wiki/Boosting_(machine_learning)
| haldujai wrote:
| This really only works well in resource limited settings
| and/or semisupervised tasks.
|
| I've tried augmentation for LLM domain adaptation and it's
| very modest gains in the best of situations, and even still
| the augmented corpus is a very tiny fraction of the
| underlying training corpus.
|
| I believe OP's question was getting at whether synthetic
| data is useful as a substantial corpus for unsupervised
| training of a language model (given the topic it's
| reasonable to disregard other areas of 'AI') and that
| answer appears to be no or at least unproven and non-
| intuitive.
| sanxiyn wrote:
| AlphaZero in fact improves based on its own output, but I
| agree it is a special case and probably not generalizable.
| Buttons840 wrote:
| It's RL though. Its output comes, in part, from interaction
| with an environment. It also has a well defined objective
| (win games). GTP doesn't have a clear objective other than
| "do more of this".
| hackerlight wrote:
| > AI can generate as much synthetic data as we need, on demand.
|
| Doesn't work in majority of domains. You need to know the
| generating process (e.g. game rules) and build a realistic
| simulation environment that emulates that, in order to generate
| data that is useful. Both of these things are out of reach for
| most applications.
|
| I believe the next large step will be multi-modal, where text
| is contextualized by video so the LLM will be able to
| concretize what "sitting on a chair" actually means with a
| single example, without needing to see thousands of textual
| associations to infer the meaning from the text.
| qeternity wrote:
| You're just sampling from an already sampled distribution.
|
| This is not the same thing. There will still be value for fine
| tuning, but it's no substitute.
| mirekrusin wrote:
| It's not one way consumer just like humans are not. It can
| direct long term evolution of reason. For starters it can be
| used to denoise/dedup/optimise training set to be closer to
| optimum (to create smaller "copies" of itself).
|
| There are instances of things that happened (history, what
| Paris Hilton did say on 22nd of April etc, big database of
| mostly irrelevant facts) and truths (math, physics, chemistry
| etc) where AI can enhance discoveries by helping us to see
| what we have not yet realised.
|
| Both seem endless tbh but personally I'm more interested in
| latter.
| version_five wrote:
| Lots of reasons this isn't universally true - it only works if
| you know enough about the data to simulate it, and your stuck
| within some distribution + human guesses space that's not all
| encompassing.
|
| The easiest counterexample is training LLMs, how are you going
| to synthesize useful language examples if you want more. Some
| version of this is true for most applications.
| mirker wrote:
| Yeah the issue is you can generate data, but it won't be good
| data. Training over random strings won't make you learn
| language, but it's technically data.
| haldujai wrote:
| To my knowledge no SOTA model has been trained on a significant
| proportion of synthetic data, has this changed?
|
| The best examples I know of are instruction tuning sets but
| that is a minute amount of data compared to the unsupervised
| training data.
| haldujai wrote:
| I wonder if the better question is not how we get more training
| data but:
|
| If we're running out of training data with hallucinations and
| performance remaining so inadequate (per OpenAI's whitepaper) is
| an autoregressive transformer the right architecture?
|
| Perhaps ongoing work in finetuning will take these models to the
| next level but ignoring the LLM hype it really does seem like
| things have plateaued for a while now (with expected gains from
| scaling).
| visarga wrote:
| There is still an order of magnitude more organic text. Ilya
| Sutskever recently said it was still ok. After that, we got to
| use reinforcement learning (agent GPTs with tools) to generate
| and self-validate more examples.
|
| One "simple" application would be to build a full index of
| facts in the whole training corpus. Just pass each document to
| GPT and ask it to extract the facts. Then create an inverted
| index, with each fact and its references. This will allow us to
| generate a wikipedia-like corpus of exhaustive fact research.
| We can say if a fact is known or not, we can tell if it is
| settled or controversial, and if it is a preference we can tell
| what is the distribution. This has got to help with factuality
| and generate lots of text to feed the model. Basically only
| costs electricity and GPU. It nicely side-steps the problem of
| truth by simply modelling the empirical distribution in an
| explicit way. At least the model won't hallucinate outside the
| known facts.
| brazzy wrote:
| >After that, we got to use reinforcement learning (agent GPTs
| with tools) to generate and self-validate more examples.
|
| How would you "self-validate" against hallucinated facts?
|
| What makes self-validation possible are hard external rules
| that can be evaluated independently and automatically. Like
| the rules of Chess or Go.
|
| We don't have anything like that for LLMs and what people
| want to use them for.
| pizza wrote:
| RLHF seems to suggest that human feedback to tune the model
| after plain textual data pretraining is quite potent per
| sample. There might be some optimal ratio of data+model
| size:rlhf size that works quite favorably for us in getting
| hallucinations to a minimum. Furthermore there might be
| some "there" there, in the hallucinations, that has yet to
| be identified as valuable in itself. Either way it seems
| like our ability to wrangle these models is getting better
| haldujai wrote:
| > There is still an order of magnitude more organic text.
|
| Posing this as a thought experiment, agree we still have more
| data to go. That we are wondering about this suggests that
| the current approach may be inadequate, i.e. it should not
| take petabytes of data for a LLM to match the performance of
| a high school student (for the LLM = AGI folks).
|
| > One "simple" application would be to build a full index of
| facts in the whole training corpus. Just pass each document
| to GPT and ask it to extract the facts.
|
| Agree, KG+LLM is a good next step to explore and should
| address some hallucination issues (see DRAGON from Leskovec
| and Liang groups). But we're already now talking about
| architectural changes as I posited.
|
| In any case, where do we get such knowledge graphs (or index
| of facts)? Some already exist (e.g. Wiki, UMLS) and were
| created by humans but are clearly inadequate in coverage.
|
| The proposition of using GPT-like models to generate these
| (i.e. GraphGPT) seems conceptually flawed as GPT does not
| itself know if a statement is factual or not which is
| problematic even for humans.
|
| Settled vs controversial is orders of magnitude more complex,
| how on earth do we do this without human annotation? You
| can't rely on frequency (i.e. some things were facts for 100
| years but all of a sudden they're not anymore and this is not
| controversial by definition).
|
| The only reason LLMs work as well as they do now is because
| sheer volume of data (and NTP) makes the noise seem hidden
| and by definition an autoregressive model should be somewhat
| impervious to singular factoids (vs a model being grounded by
| the garbage dump that is CommonCrawl/the internet).
|
| > At least the model won't hallucinate outside the known
| facts.
|
| Not sure this is a given, even if a model acts as a natural
| language database of factoids it is probable that it will
| hallucinate links unless you're strictly grounding output in
| which case we've just built a colossally over-engineered
| IR/STS tool.
|
| > One "simple" application
|
| I think what you've posited is actually harder to build than
| anything that's been achieved thus far with LLMs.
| brianr wrote:
| This analysis misses the impact of AI models being deployed, like
| is happening rapidly right now. Production applications built on
| AI will provide ample (infinite?) additional training data to
| feed back into the underlying models.
| haldujai wrote:
| Not sure that synthetic or LLM-generated training data is as
| useful as human generated text.
|
| It seems "good enough" (for now) but synthetic makes up a very
| small proportion of the training set being used in current
| models that have been trained on it, if that proportion ends up
| being mostly synthetic we'll likely see whatever weird
| hallucinations and biases in the dominant backend (GPT4 or
| whatever) become amplified.
|
| It's been shown repeatedly that garbage in = garbage out for
| training data.
| replygirl wrote:
| If we play our cards right, AI could free people up for more
| valuable pursuits, and the pace of human information production
| would increase by orders of magnitude
| visarga wrote:
| > free people up for more valuable pursuits
|
| It won't roll like that. AI will empower people to be more
| productive but won't free people up because it makes mistakes,
| can't help itself, and cannot function autonomously. There is
| no LLM application that is safe for autonomous usage today. How
| can we go from 0 to 1? I don't see a path. Self driving cars
| still can't reach L5 to completely remove the need for driver.
|
| But maybe this is a blessing in disguise. It will make AI more
| like a new ability of humans than of the companies. Companies
| need people to unlock AI efficiencies. And AI tends to become
| open sourced so everyone has access to the same. AI is not a
| moat for companies and human ability to hand-held it is tied to
| individuals. That would make the transition easier. Solving
| that last 1% accuracy might encounter exponential friction and
| last for a while.
| replygirl wrote:
| Yeah, it's not that I think we'll get all the way there, it's
| a utopia. My expectation is that within 30 years we reduce
| the work week by a day or two for most people, compensate for
| our education system's decline, and avoid energy and food
| crises, and nothing else fundamentally changes
| ChatGTP wrote:
| I can't see us going from Microsoft and Open AI stealing
| everyone's work and selling it's without attribution or
| respect for GPL (for example ) to technological utopia
| anytime soon.
| jacooper wrote:
| Assuming this happens without any violence, which I truly
| doubt. Huge socioeconomic changes like this always come as
| a result of violence and uproar.
| gumballindie wrote:
| > free people
|
| As opposed to what? Being "captive" in jobs for paying bills?
| digdugdirk wrote:
| I mean... Yes.
|
| What would you suggest as the alternative?
| gumballindie wrote:
| Erm not causing mass unemployment by stealing data? Also
| people go freely where there's pay. Seems like there aren't
| many opportunities and there will fewer.
| replygirl wrote:
| So hopefully we play our cards right, by extracting
| benefits from AI that overcompensate for the negative
| impacts like mass unemployment and democratization of
| intellectual property. If the spoils are distributed in
| such a way that people's standard of living is maintained
| or improved, people have more liesure time, which the
| social sciences have shown will not mean people will just
| stop working--they'll work less, but with higher
| productivity on things promising a greater benefit to
| family, community, and society.
|
| Forgive me if I'm misreading, but I'm having trouble with
| your line of reasoning. Your first reply to me
| scarequoting "captive" strongly implies an argument that
| the imperative to seek employment for survival is not a
| limiting factor on how people spend their time, and
| therefore that my suggestion that giving people more
| choice over how they apply their talents could be a good
| thing is irrelevant; but your child reply implies a
| concern that AI taking over some human labor will cause
| mass unemployment and explicitly states choice is
| declining.
|
| I'm advocating that, since the genie is out of the
| bottle, AI could be used to free people from toil, just
| as other labor innovations like machinery and the 40-hour
| work week have done. Why the dismissive snark? In the
| abstract, do we not want the same thing?
| gumballindie wrote:
| Unfortunately real life doesn't work the way you
| describe. People won't be "free" to enjoy more "leisure"
| time, by "democratizing" the results of their work.
| Instead, all of this stolen data, will be used to "free"
| them from jobs and to consolidate corporate control. Say
| bye to microbusinesses, to freelancers, to indie
| developers. You know, those people that have been truly
| free. Similarly, office workers will be "free" to lose
| the jobs that they chose to perform, and will have
| nowhere else to go but unemployment lines. All thanks to
| "democratizing" ip by stealing data.
| replygirl wrote:
| I agree that not playing our cards right is the default
| scenario.
| gumballindie wrote:
| As a technology ai can indeed free people in a productive
| manner. But it would appear that we started on the wrong
| footing. Power will be consolidated in the hands of a few
| at a scale we haven't seen yet. It all depends tho on
| whether we can regulate how data is collected, at least
| at the basic level of not infringing copyright.
| visarga wrote:
| > All thanks to "democratizing" ip by stealing data.
|
| That horse is already dead. Large models can learn
| everything, there's nothing that can be done to stop them
| from learning. It's too easy for them to do it. We can't
| hold any meaningful IP when models can generate 100
| variations only different enough to pass the test. IP is
| dead. But on its corpse there will grow a new world of
| applications. We all got new skills, depends on us if we
| use them or not.
| gumballindie wrote:
| Sure. In before people used to say that currency is dead
| because crypto currency has replaced money already, and
| already people are using them, and already [insert
| marketing statement]. I see they now moved on to ai.
| mkaic wrote:
| I agree! I actually wrote a blogpost about this
| recently[0], but the TLDR is that ownership is nothing
| without enforcement, and it has become increasingly
| difficult to enforce ownership of intellectual property
| in the modern world -- first digital files, then the
| sharing of those digital files over the Internet, and now
| generative models that allow people to create high-
| quality ripoffs of _any_ IP for zero marginal cost. The
| sheer volume will just be too much to contend with,
| because you can 't sue _everyone_. In my view, this is a
| good thing and a long time coming!
|
| [0] https://blog.kaichristensen.com/p/generative-ai-is-
| the-final...
| blibble wrote:
| > the pace of human information production would increase by
| orders of magnitude
|
| you mean boilerplate and spam right?
| nologic01 wrote:
| Brute force approaches always hit some wall. ML will be no
| different. In the decades to come it us quite likely that
| algorithms will develop in directions orthogonal to current
| approaches. The idea that you improve performance by throwing
| gazillions of data into gargantuan models might be even come to
| be seen as laughable.
|
| Keep in mind (pun) that the only real intelligence here is us,
| and we are pretty good at figuring out when a tool has exhausted
| its utility.
| gleenn wrote:
| AI had a winter of many decades because the hardware wasn't
| there and there were better alternatives, especially for neural
| nets. Now ChatGPT etc comes out, with unbelievable results,
| decades in the making. And a couple months we're already
| writing it off because of the next limitation? Maybe let's give
| it more than a month or two to figure out if we even need all
| that data. I heard they're already talking about trying to
| significantly reduce the model hyper parameters size even
| though a large model size increase apparently the reason GPT 4
| was so much better than 3. Give it a minute IMHO before making
| generalizations like this so soon
___________________________________________________________________
(page generated 2023-04-22 23:01 UTC)