A variational approximate posterior for the deep Wishart process
Authors: Sebastian Ober, Laurence Aitchison
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | For more quantitative experiments, we trained a DWP and a DGP with the exact same generative model with a squared exponential kernel on the UCI datasets from Gal & Ghahramani (2015). We trained both models for 20000 gradient steps using the Adam optimizer Kingma & Ba (2015); we detail the exact experimental setup in Appendix F. We report ELBOs and test log likelihoods for depth 5 in Table 1 |
| Researcher Affiliation | Academia | Sebastian W. Ober Department of Engineering University of Cambridge Cambridge, UK swo25@cam.ac.uk Laurence Aitchison Department of Computer Science University of Bristol Bristol, UK laurence.aitchison@bristol.ac.uk |
| Pseudocode | Yes | Algorithm 1 Computing predictions/ELBO for one batch |
| Open Source Code | Yes | We provide a reference implementation at https://github.com/Laurence A/bayesfunc. |
| Open Datasets | Yes | on the UCI datasets from Gal & Ghahramani (2015). |
| Dataset Splits | Yes | For more quantitative experiments, we trained a DWP and a DGP with the exact same generative model with a squared exponential kernel on the UCI datasets from Gal & Ghahramani (2015). We detail the exact experimental setup in Appendix F. We show a plot of the training curves for one split of boston with a 5-layer DGP and DWP in Fig. 1 |
| Hardware Specification | No | In Appendix H, we provide a table of time per epoch, which shows that we obtain faster runtime for protein and for shallower models, although the gains are slightly more modest due to the models being shallower and the fact that we run protein on a GPU, as opposed to a CPU for boston. |
| Software Dependencies | No | The paper mentions using the Adam optimizer but does not provide specific version numbers for software dependencies or frameworks. |
| Experiment Setup | Yes | We trained both models for 20000 gradient steps using the Adam optimizer Kingma & Ba (2015); we detail the exact experimental setup in Appendix F. |