analiza aiqbench
This arXiv preprint offers a theoretical account of how diffusion models can memorize individual training samples instead of learning the underlying data distribution, and proposes a criterion meant to flag that failure. Its claims rest on analytical characterization and experiments across architectures designed to both promote and suppress memorization, so the practical value depends on how well the criterion holds up outside those settings. Readers working on generative model evaluation, privacy, or training diagnostics should check whether the detector transfers to larger, production-scale models.
Detection of Memorization in Diffusion Models Through Thermodynamic Singularity
arXiv:2610.27754v1 Announce Type: new Abstract: In diffusion models, proper distributional learning entails reversing a diffusion process over underlying density functions, whereas a key failure is memorization, in which the reverse process instead learns diffusion from individual training samples. We provide an analytical characterization of such memorized reverse processes, along with a criterion that detects the associated thermodynamic singularity. The proposed criterion reliably distinguishes between properly learned and memorizing networks across a range of neural network models designed to promote or mitigate memorization.
Abstract:In diffusion models, proper distributional learning entails reversing a diffusion process over underlying density functions, whereas a key failure is memorization, in which the reverse process instead learns diffusion from individual training samples. We provide an analytical characterization of such memorized reverse processes, along with a criterion that detects the associated thermodynamic singularity. The proposed criterion reliably distinguishes between properly learned and memorizing networks across a range of neural network models designed to promote or mitigate memorization.
Machine Learning (cs.LG)
arXiv:2610.27754 [cs.LG]
(or arXiv:2610.27754v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.27754 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 171 cuvinte
Subiecte conexe
Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
Latent Space · 8 oct. 2026
Thermodynamic Weight Decay: Exploring Grokking Acceleration via Attention Specific Heat
arXiv · 24 iul. 2026
Towards Automated Discovery: A Review of Generative Models, Multimodal Learning and Closed-Loop Workflows in Inverse Materials Design
arXiv · 24 aug. 2026
Lunch with Horace Walpole
Quantum Frontiers · 26 oct. 2026