Latent Abstractions in Generative Diffusion Models

In this work, we study how diffusion-based generative models produce high-dimensional data, such as images, by relying on latent abstractions that guide the generative process. We introduce a novel theoretical framework extending Nonlinear Filtering (NLF), offering a new perspective on SDE-based gen...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Entropy (Basel, Switzerland) Jg. 27; H. 4; S. 371
Hauptverfasser: Franzese, Giulio, Martini, Mattia, Corallo, Giulio, Papotti, Paolo, Michiardi, Pietro
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Switzerland MDPI AG 31.03.2025
MDPI
Schlagworte:
Analysis
Brownian motion
Computational linguistics
Diffusion models
Information theory
Language processing
Natural language interfaces
nonlinear filtering
Random variables
world modeling
Germany
Michigan
information theory
nonlinear filtering
diffusion models
world modeling
ISSN:1099-4300, 1099-4300
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In this work, we study how diffusion-based generative models produce high-dimensional data, such as images, by relying on latent abstractions that guide the generative process. We introduce a novel theoretical framework extending Nonlinear Filtering (NLF), offering a new perspective on SDE-based generative models. Our theory is based on a new formulation of joint (state and measurement) dynamics and an information-theoretic measure of state influence on the measurement process. We show that diffusion models can be interpreted as a system of SDE, describing a non-linear filter where unobservable latent abstractions steer the dynamics of an observable measurement process. Additionally, we present an empirical study validating our theory and supporting previous findings on the emergence of latent abstractions at different generative stages.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
content type line 23
ISSN:1099-4300
1099-4300
DOI:10.3390/e27040371