Score-Based Neural Processes

26 Sept 2024 (modified: 19 Dec 2024)ICLR 2025 Conference Withdrawn SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Neural Processes, Diffusion Models, Generative Models, Stochastic Processes
TL;DR: We propose a method for incorporating score-based generative models into the neural process paradigm
Abstract:

Neural processes (NP) are a flexible class of models that generate stochastic processes by operating on finite-dimensional marginal distributions. NPs are designed to maintain exchangeability and marginal consistency, which are necessary to define a valid stochastic process. However, NP variants can come with drawbacks such as limited expressivity, uncorrelated samples, and consistency sacrifices. To address the issues of previous NPs, we introduce score-based neural processes, \emph{scoreNP}, which incorporate a score-based generative model within the neural process paradigm. This score-based approach enhances expressivity, allowing the model to capture complex non-Gaussian distributions of functions, generate correlated samples, and maintain marginal consistency. Previously, no NP variant has been able to maintain conditional consistency. We show that using \emph{guidance} methods from conditional diffusion sampling, \emph{scoreNP} is the first NP is able to satisfy conditional consistency. Empirically, \emph{scoreNP} performs well qualitatively and quantitatively well across a range of unconditional and conditional functional generation tasks.

Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8267
Loading