Learning Nonparametric Volterra Kernels with Gaussian Processes
Authors: Magnus Ross, Michael T Smith, Mauricio Álvarez
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the performance of the model for both multiple output regression and system identification using standard benchmarks. 5 Experiments For all the following experiments we place the inducing locations for both the input process and VKs on a fixed grid. |
| Researcher Affiliation | Academia | Magnus Ross Department of Computer Science University of Sheffield, UK mross1@sheffield.ac.uk Michael T. Smith Department of Computer Science University of Sheffield, UK m.t.smith@sheffield.ac.uk Mauricio A. Álvarez Department of Computer Science University of Sheffield, UK mauricio.alvarez@sheffield.ac.uk |
| Pseudocode | No | The paper describes methods textually and with mathematical equations, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at github.com/magnusross/nvkm. |
| Open Datasets | Yes | To demonstrate the IO-NVKM, we use a standard benchmark for nonlinear systems identification know as Cascaded Tanks [24].3 The system comprises two vertically stacked tanks... 3Available at sites.google.com/view/nonlinear-benchmark/ and To illustrate the utility of the NVKM for multiple output regression problems, we consider a popular benchmark in the MOGP literature, consisting of multiple correlated time series of air temperature measurements taken at four nearby locations on the south coast of England, originally described by Nguyen et al. [20], which we refer to as Weather.4 4Available for download in a convenient from using the wbml package, github.com/wesselb/wbml |
| Dataset Splits | No | The paper describes training and testing splits for the datasets (e.g., 'a random subset of a third for training and the rest for testing'), but does not explicitly provide details for a separate validation dataset split. |
| Hardware Specification | Yes | All models were trained on a single Nvidia K80 GPU. |
| Software Dependencies | No | The paper mentions software used ('Jax framework', 'Adam') but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | For all the following experiments we place the inducing locations for both the input process and VKs on a fixed grid. ... For all experiments we use 15, 102 = 100, 63 = 216 and 44 = 256 inducing points, for each of the 1st to 4th order filters respectively, centered on zero. We treat the range of the points of each VK as a hyperparameter, and fix α such that the decaying part of the DSE covariance causes samples to be near zero at the edge of the range. ... The VK GP length scales, VK GP amplitudes, and input GP amplitude are optimised along with the variational parameters by maximising the variational bound using gradient descent. ... the noise hyperparameters are fixed to a small value whilst the other hyperparameters and variational parameters are optimised, and then estimated afterwards by minimising the bound with all other variables fixed. |