Input Dependent Sparse Gaussian Processes

Authors: Bahram Jafrasteh, Carlos Villacampa-Calvo, Daniel Hernandez-Lobato

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method in several experiments, showing that it performs similar or better than other stateof-the-art sparse variational GPs. However, in our method the number of inducing points is reduced drastically since they depend on the input data. This makes our method scale to larger datasets and have faster training and prediction times.
Researcher Affiliation Academia 1Biomedical Research and Innovation Institute of C adiz (INi BICA) Research Unit, Puerta del Mar University, C adiz, Spain 2Computer Science Department, Universidad Aut onoma de Madrid, Madrid, Spain.
Pseudocode Yes Algorithm 1 Training input dependent sparse GPs
Open Source Code Yes The code of IDSGP in Tensorflow 2.0 (Abadi et al., 2015) is given in the supplementary material.
Open Datasets Yes All the datasets are publicly available. The UCI repository datasets can be downloaded from the repository (Dua & Graff, 2017). Yellow taxi dataset was preprocessed following Salimbeni & Deisenroth (2017) and downloaded from https: //www1.nyc.gov/site/tlc/about/tlc-trip-record-data.page, where we have used data records from year 2015. Similarly, the Airlines Delay dataset was preprocessed following Hern andez-Lobato & Hern andez-Lobato (2016) and was downloaded from https://community.amstat.org/jointscsg-section/dataexpo/ dataexpo2009, keeping only the records from January 2008 to April 2008.
Dataset Splits No The paper states: "On each dataset we use 80% of the data for training and the rest for testing." It does not explicitly mention a separate validation split or how it was handled if used (e.g., cross-validation, fixed percentage).
Hardware Specification Yes All methods are trained on a Tesla P100 GPU with 16GB of memory.
Software Dependencies Yes The code of IDSGP in Tensorflow 2.0 (Abadi et al., 2015) is given in the supplementary material.
Experiment Setup Yes All the methods are trained using ADAM (Kingma & Ba, 2015) with a mini-batch size of 100 and a learning rate of 0.01. In the classification setting we use the same setup, but the number of inducing points of IDSGP is set even smaller. Namely, M = 3. All methods are trained on a Tesla P100 GPU with 16GB of memory. On each dataset we use 80% of the data for training and the rest for testing. We report results across 5 splits of the data since the datasets are already quite big. The DNN architecture used in IDSGP is detailed in Appendix B.