Скачать или смотреть AI models collapse when trained on recursively generated data

Recent foundation models increasingly rely on large-scale synthetic data for training, enabling scalable supervision but also introducing a fundamental failure mode known as model collapse, where recursive training on model-generated data leads to progressive loss of diversity, elimination of rare patterns, and convergence toward dominant modes. In this seminar, I will examine model collapse from theoretical and empirical perspectives, showing that it arises even under idealized assumptions such as exact functional approximation, due to finite sampling effects that can be characterized as an absorbing stochastic process. I will review experimental evidence in language models demonstrating how diversity reduction precedes performance degradation, discuss recent findings on strong model collapse that reveal how even small fractions of synthetic data can disrupt scaling laws—especially in large models—and conclude by surveying mitigation strategies such as real-data accumulation, synthetic data verification, and data provenance management, as well as emerging perspectives that leverage controlled model collapse for applications like machine unlearning and model editing.

Presenter: Jaeik Kim (Shumailov et al., Nature 2024)
Paper: AI models collapse when trained on recursively generated data