Self-Consuming Generative Models Go MAD
arxiv.org
arxiv.org
> with enough fresh real data, the quality and diversity of the generative models do not degrade over generations
When a model gets deployed and is prompted by people it can get fresh data in the prompt - the prompt itself, RAG material, outputs of tools, human responses to its outputs - and this allows for a form of exploration that can go outside the original scope of the model. If you use the model logs to retrain it won't collapse.
AI is only as good as its models. Now that we're flooding the internet with AI generated content with no way to determine if the content was also produced by AI while endlessly shoveling as much content as possible into these AI models to push up those parameter counts to attract investors, it's inevitable the current generation of popular AI systems will just eat themselves.
If they also use RAG or other tools like code execution there is even more ample chance to steer the model outside its original experience.
Then we can use these sessions as training data to re-tune the base model. It's much cheaper than training a new model, can be done on normal computers.
LLMs are not beholden to their training set. They can ingest new information any time, or relate two things that were previously separate to derive new insights.
I find that catchy and descriptive on more than one level.
If I give a pig something to eat and he throws up I'm surely not touching that stuff myself
Whether it's an AI eating its own training data, a human stuck in an empty cell and forced to live with just his/her own thoughts, or a culture believing its own myths and aggressively punishing insiders and outsiders who challenge them - it's all solitary confinement. And we know how that goes.
(/s of course - although the self-referentiality of memes is balanced by the mad scramble for novelty.)
Or to put it in your analogy, If I feed my cat milk, they'll throw up. That doesn't mean milk is unfit for human consumption.
If AI generated stuff is so bad you cannot train AI with it why would humans use it to train themselves (aka learn stuff)?
a more relevant analogy for llms is a bullshiter that don't know shit about anything but like talk about everything and they're good at that (talking like they know), I'm sure we all at least know or knew in the past someone like that.
at a low enough percentage these bullshiters can thrive and get praise from other people but when there a lot of them and not enough of the real deal, well everyone is in trouble including the bullshiters since they no longer have who to mimic.
This space is where the Venn diagram of marketing and genocides overlap.
At least this is what our specilist at the hospital explained to us.
Obviously ethically we probably should not normalize eating people though.
It's not unqualified bad. It is bad if you do it for many times in a loop, and without any external inputs. But you can put fresh material right in the prompt and it won't suffer from retraining on synthetic data.
If you train a small model from scratch on purely synthetic data, like the Microsoft Phi models, it comes out competent and 5x more efficient than models trained on web text. So there's the flip side. You can do it in bad way, and you can do it in a good way.
A pig (or AI) throwing up on an input should increase your priors that it may be bad for humans, but it does not prove the matter either way.
If you interact with people / fresh data regularly you'll avoid that.
I'm not sure how much is down to evolving requirements and data formats vs degradation of the training data. The idea sounds common-sense but it's also triggering a bullshit warning for me, because I'd expect more consistency per AI model, and differences to occur on new model interations - regardless of the dataset.
Just because it doesn't work with AI now, doesn't mean it's something that wouldn't ever work.
AlphaZero/MuZero game AIs do very well feeding themselves. Often better than previous AIs with hardcoded rules and training based on human games.
Training AI on AI output is less like organised education systems and the less formal written & oral traditions that predate them, and more like dark reddit/chan/other communities eating each other's rhetoric, or some political¹ and religious groups in the wider world. These relatively insular (but sometimes large) groups often descend into a pit they have difficulty reasoning themselves out of.
Maybe the answer for AI is to work out what sort of external mixing helps to keep humans on track, and if it can be emulated in the training models used, so their reasoning continues to grow instead of falling into this pothole.
This might be a crap idea (too little sleep, caffeine has just made me tired and jittery) but perhaps the scale of the information we pile into the process and the fact the retraining at least partly on AI output seems inevitable, means we need to look at training an AI more like training a small population of humans rather than a single unit human.
----
[1] mostly the far right, but that might in part be because they tend to be loud so noticed – far anything is a problem
There is a difference to talking only to yourself and talking with other people.
We know that there is more than one AI software but there is not as much variation as for humans.
And as someone else already wrote here, humans have something like a ratcheting mechanism. I would paraphrase it as "Humans have learnt to stand on shoulders of giants". AI does not have this.
Luckily we don't hear too much about Mad Cow Disease these days. Let's hope it doesn't come back.
At least AI hasn't reinvented Mutually Assured Destruction!
It is not at all necessarily surprising, I think, from a purely high-level perspective, but I do personally think that I find that it is good to have the analysis. From a purely professional standpoint, I do not believe it is unique or distinctive enough as an individual method to need its own separate name for day-to-day use. From a personal perspective, however, I thought the mad cow disease reference was hilarious and applaud whoever came up with the acronym.
I find the benefit in the analysis, and the concerns presented about generated data being present in the data makes sense to me (and if in sufficient quantity, would make sense as biasing the models improperly in a rather significant kind of way).
I particularly enjoyed the humor of this line, the tongue-in-cheek nature is very funny/nice to me here:
"Ascertaining whether an autophagous loop has gone MAD or not (recall Definition 2.1) requires that we measure how far the synthesized data distribution Gt has drifted from the true data distribution Pr over the generations t."
I like their use of color in the paper, I saw a similar orange/green color scheme earlier today and enjoyed it very much as an annotation method.
"A fixed real dataset only slows generative model degradation" is again also a natural consequence of Shannon's noisy channel capacity theorem, one can say that with almost nearly perfect certainty that a limited neural network will not be able to perfectly fit the distribution of the data that it is training on, thus it will have bias, variance, or some combination of both, limited ultimately by the model's capacity itself.
This w.r.t. the original dataset is noise, and we can choose between whether we want collapse, or recursively encoding the noise patterns of the previous model (which might happen to have an additive effect, or maybe not! Who knows! I do not know for sure here, I have not yet figured this one out myself yet).
w.r.t. the real data slowing down degradation, if we are sampling I.I.D. of course then proportionately we still should see some degradation as this is the nature of empirical risk minimization over maximum likelihood estimation. It is still good that they have shown this, however, I thinks.
The fresh data loop, I believe, would be an example of actually a kind of noise in and of itself, w.r.t. the original input dataset, and as long as this 'noise' (from the perspective of the model) has a higher SNR than the (potentially slow) collapse of the model's output distribution, then it should (in some kind of proportion at leasts) be constantly-playing 'keep-up' with the fresh data.
"First, we find that—regardless of the performance of early generations—the performance of later generations converges to a point that depends only on the amounts of real and synthetic data in the training loop. " -- there we are (I saw this after making the SNR point, this makes sense within this framework of interpretation, then.
All in all, I found this paper very aware of itself and what it was studying, it was well-laid out and accessible, and while the points are not necessarily earth-shattering (though I still have to read through some of it, I think), having clear empirical evidence about this phenomenon, detailing it, and cutting away through the forest of (at-least-seemingly) untested battlegrounds is one that I appreciated.
Curious to hear what others think about this one. <3 :'))))