Reproduce “Structured Bayesian Gaussian process latent variable model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Steven Atkinson, Nicholas Zabaras
The Structured Bayesian Gaussian Process Latent Variable Model (SGPLVM) from the University of Notre Dame introduces a method for learning low-dimensional representations of high-dimensional, spatially correlated data. It achieves linear computational complexity by exploiting Kronecker product structures, demonstrating improved data imputation and novel super-resolution capabilities on image and video datasets.
Key insights
- Exploiting Kronecker product structures in GP kernels enables linear computational complexity, making Bayesian GP-LVMs scalable for high-dimensional structured data.
- Explicitly modeling spatial correlations within observations via a parameterized spatial kernel improves sample efficiency and allows for predictions at varying spatial resolutions.
- A "partially-collapsed" variational bound provides an efficient mechanism for inferring latent variables and performing data imputation on partially-observed test cases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.