Beyond the Mean-Field: Structured Deep Gaussian Processes Improve the Predictive Uncertainties
Authors: Jakob Lindinger, David Reeb, Christoph Lippert, Barbara Rakitsch
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply it to several benchmark datasets. It yields excellent results and strikes a better balance between accuracy and calibrated uncertainty estimates than its state-of-the-art alternatives. In Sec. 4, we show experimentally that the new algorithm works well in practice. |
| Researcher Affiliation | Collaboration | 1Bosch Center for Artificial Intelligence, Renningen, Germany 2Hasso Plattner Institute, Potsdam, Germany 3University of Potsdam, Germany {jakob.lindinger, david.reeb, barbara.rakitsch}@de.bosch.com, christoph.lippert@hpi.de |
| Pseudocode | Yes | A pseudocode description of our algorithm is given in Appx. F. |
| Open Source Code | Yes | Python code (building on code for the mean-field DGP [25], GPflow [19] and Tensor Flow [1]) implementing our method is provided at https://github.com/boschresearch/Structured_DGP. |
| Open Datasets | Yes | We report results on eight UCI datasets and employ as evaluation criterion the average marginal test log-likelihood (tll). |
| Dataset Splits | Yes | We assessed the interpolation behaviour of the different approaches by randomly partitioning the data into a training and a test set with a 90 : 10 split. To investigate the extrapolation behaviour, we created test instances that are distant from the training samples: ...divided them accordingly into training and test set using a 50 : 50 split. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments, such as GPU/CPU models, memory specifications, or cloud computing instance types. |
| Software Dependencies | No | The paper mentions software like 'GPflow' and 'Tensor Flow' but does not specify their version numbers, which is required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | We used a fully-coupled DGP with our standard three layer architecture (see Sec. 3.2), on the concrete UCI dataset trained with Adam [15]. For our standard setting, M = 128, our STAR approximation was only two times slower than the mean-field but three times faster than FC DGP (trained with Adam [15]). |