This is a text automatically translated from Italian. If you appreciate our work and if you like reading it in your language, consider a donation to allow us to continue doing it and improving it.
The articles of Cassandra Crossing I'm under license CC BY-SA 4.0 | Cassandra Crossing is a column created by Marco Calamari with the "nom de plume" of Cassandra, born in 2005.
Cassandra returns (here is the first part) on AI with always valuable advice!
This article was written on February 29, 2024 from Cassandra
Cassandra Crossing 575/ AI: Collapse, reloaded
Why should a technical detail of false AI, even a mathematical one, be known to everyone?
Those of the 24 indefatigable readers who have already tried their hand at reading Cassandra's previous statement on the same topic can safely continue; everyone else should just read it, as if it were the first half of a mystery film to which you arrived late.
The second half begins, whose screenwriter is not Cassandra, but a deserving person who posted on Reddit a well-made summary of a rather abstruse question, described in a specialist paper.
Disclosure: Cassandra reworked it, translated it (with an LLM) and then reworked it again, and inserts it here. An obligatory thank you to the real author, then, and let's move on.
Warning: a minimum amount of mathematics is necessary, but if you have ever taken an exam or a thesis on basic statistics, it will be very clear to you, indeed dazzling in its simplicity.
. . .
The collapse of a modelor LLM occurs when a model, trained on data generated by previous generations of models, begins to lose information, particularly at the tails of the statistical distribution of the original data, and eventually converges to a single-point estimate, with little statistical variance. (Editor's note: in a nutshell, he always gives the same answer to any question).
The model collapse occurs due to two sources of error: statistical approximation error due to finite sampling, and functional approximation error due to imperfect models. These errors accumulate over generations, causing the estimated distribution to deviate further from the original.
Some studies show that model collapse occurs in simple models such as i Gaussian Mixture Models and of Variational Autoencoders, as well as in Linguistic models more complex. Even a model calibration during training does not prevent the model collapse in linguistic models.
Over time, the data generated by the affected models model collapse they begin to contain unlikely sequences, and lose information about the statistical tails of the original distribution.
The researchers argue that, to avoid model collapse and maintain its performance, access to data generated directly by humans, to be used for training, will remain essential. Furthermore, the data produced during human interactions with language models will be increasingly valuable.
In summary, we highlight an important phenomenon in which models trained recursively on their generated data begin to lose fidelity, and eventually converge towards a suboptimal state.
Access to original human-generated data is therefore necessary to avoid model collapse, and sustain its performance in the long term. For this reason, as language models begin to generate an increasing amount of the web's content, distinguishing human data from model-generated data on a large scale will become a major challenge in successfully training language models.
. . .
Cassandra, really doing her job this time, ends with a warning.
Yes, let's talk about them false artificial intelligence which you can already find in the latest versions of the products you use. Even in computer code generators, such as Copilot, which most developers now use. And therefore whose effects you can find in the software you use and undergo every day, and who will check the next version of the intelligent objects.
Marco Calamari
Video column “A chat with Cassandra”
Cassandra's Slog (Static Blog).
Cassandra's archive: school, training and thought
This tag @loyal alternatives is used to automatically send this post to Feddit and allow anyone on the fediverse to comment on it.
Join communities
If you have found errors in the article you can report them by clicking here, Thank you!
Comments
Each article corresponds to a post on Feddit where you can comment! ✍️ Click here to comment on this article ✍️
Feddit is the Italian alternative to Reddit managed by us, based on the software Lemmy, one of the most interesting projects of fediverse.
1 comment
Puoi commentare questo articolo dal tuo account sul Fediverso! Se hai Mastodon o Lemmy (Feddit.it) cerca il link di questo articolo e commenta dal Fediverso! .
It's not to break the eggs in the basket but, from blogger old school, for example in this passage:
I would probably have put at least 2 or 3 link to the paper used to write this passage. Then perhaps there is a third source who has already examined and mulled over the matter, only this third source is also missing. And then, although it is written about warning initial that
and indeed it is, but this does not mean that I have to start looking for the sources and retrace the same research as the author, who then in the end have enough material to write a book myself. post. Or, from the point of view of a reader not in the "profession" (in the sense of scientific communication), what is written is taken as "cast gold".