Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Sparse Gaussian Processes: Structured Approximations and Power-EP Revisited
Authors: Thang Bui, Michalis K. Titsias
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive regression experiments, we show that the proposed block-diagonal approximation consistently performs similarly to or better than existing diagonal approximations while maintaining comparable computational costs. Furthermore, the new PEP framework with structured posteriors provides competitive performance across various power hyperparameter settings, offering practitioners flexible alternatives to standard variational approaches. |
| Researcher Affiliation | Collaboration | Thang D. Bui Australian National University EMAIL Michalis K. Titsias Google Deep Mind EMAIL |
| Pseudocode | No | The paper describes the methods verbally and mathematically, including detailed derivations in the main text and appendix, but does not present a structured pseudocode or algorithm block. |
| Open Source Code | Yes | We provide an implementation here https://github.com/thangbui/tighter_sparse_gp. |
| Open Datasets | Yes | We next ran an experiment to validate the utility of the proposed block-structured approximation in section 3 on four real-world regression datasets1. For each dataset and each inducing point configuration (M = 256 or M = 512), we compare the uncollapsed variational bounds... We used the splits available in this repository https://github.com/treforevans/uci_datasets. |
| Dataset Splits | Yes | We repeated the experiment 10 times, each using a random train/test split, a batch size of 500 (also the block size), random partitioning of the training data into blocks, and 300 epochs for training. |
| Hardware Specification | Yes | All experiments were done on either a V100 GPU or a Mac Book laptop. |
| Software Dependencies | No | The implementation was built on GPflow, and released here https://github.com/thangbui/tighter_sparse_gp. |
| Experiment Setup | Yes | We used 5 inducing points in this experiment. ... We repeated the experiment 10 times, each using a random train/test split, a batch size of 500 (also the block size), random partitioning of the training data into blocks, and 300 epochs for training. ... For the later datasets, we used the median distance between the data points to initialise the lengthscales and set the initial observation noise variance to 0.1. ... In the Snelson, kin40k, and Power-EP experiments, we optimised the collapsed bound using the L-BFGS optimiser. In the block-diagonal experiments with medium-scale datasets, we used the Adam optimiser with a learning rate of 0.005. |