Lossy Loops: Shannon's DPI and Information Decay in Generative Model Training

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: Straňák, Pavel
Natura: Recurso digital
Lingua:inglese
Pubblicazione: Zenodo 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866902000702062592
author Straňák, Pavel
author_facet Straňák, Pavel
contents <p>Model collapse, the progressive degradation of generative AI performance when trained on synthetic data, poses a critical challenge for modern AI systems. This paper establishes a theoretical framework based on Shannon's Data Processing Inequality (DPI) to explain this phenomenon. We conceptualize generative AI models as lossy communication channels, predicting progressive mutual information decay during iterative training. We derive testable hypotheses for exponential decay rates (λ ∈ [0.2, 0.4] per iteration) and propose mitigation paradigms requiring future validation.</p> <p>See also:<a href="https://doi.org/10.5281/zenodo.15199262"> https://doi.org/10.5281/zenodo.15199262</a> for a related philosophical exploration.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_16481493
institution Zenodo
language eng
publishDate 2025
publisher Zenodo
record_format zenodo
spellingShingle Lossy Loops: Shannon's DPI and Information Decay in Generative Model Training
Straňák, Pavel
Model collapse, Information theory, Data Processing Inequality, Generative AI, Synthetic data, Mutual information, AI limitations
<p>Model collapse, the progressive degradation of generative AI performance when trained on synthetic data, poses a critical challenge for modern AI systems. This paper establishes a theoretical framework based on Shannon's Data Processing Inequality (DPI) to explain this phenomenon. We conceptualize generative AI models as lossy communication channels, predicting progressive mutual information decay during iterative training. We derive testable hypotheses for exponential decay rates (λ ∈ [0.2, 0.4] per iteration) and propose mitigation paradigms requiring future validation.</p> <p>See also:<a href="https://doi.org/10.5281/zenodo.15199262"> https://doi.org/10.5281/zenodo.15199262</a> for a related philosophical exploration.</p>
title Lossy Loops: Shannon's DPI and Information Decay in Generative Model Training
topic Model collapse, Information theory, Data Processing Inequality, Generative AI, Synthetic data, Mutual information, AI limitations
url https://doi.org/10.5281/zenodo.16481493