Measuring generative model decay under iterative retraining with its own data

Measuring generative model decay under iterative retraining with its own data

Computer Science

David Litchfield

2024

Ranko Lazic and Matthias Englert

This research project explores the rise of AI-generated content, termed “synthetic content,” and its implications for generative models. Prior research indicates that if the ratio of synthetic to real content is too high, models may experience “collapse,” resulting in a decline in output quality. This study aims to analyse how increasing the synthetic ratio affects model behaviour, investigating whether this leads to a gradual decline or a sudden collapse. Understanding this boundary could help online service providers regulate synthetic content and improve insights into the behaviour of generative AIs.

research,poster,machine learning,generative ai