Near-linear time Gaussian process optimization with adaptive batching and resparsification

Authors: Daniele Calandriello, Luigi Carratino, Alessandro Lazaric, Michal Valko, Lorenzo Rosasco

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental These findings are then confirmed in several experiments, where BBKB is much faster than state-of-the-art methods.
Researcher Affiliation Collaboration 1Istituto Italiano di Tecnologia, Genova, Italy (now at Deepmind, Paris, France) 2Ma LGa Dibris Universit a degli Studi di Genova, Italy 3Facebook AI Research, Paris, France 4Deep Mind, Paris, France 5MIT, Cambridge, MA, USA 6Istituto Italiano di Tecnologia, Genova, Italy.
Pseudocode Yes Algorithm 1 BBKB
Open Source Code Yes Code can be found at github.com/luigicarratino/batch-bkb
Open Datasets Yes We first perform experiments on two regression datasets Abalone (A = 4177, d = 8) and Cadata (A = 20640, d = 8) datasets. We then perform experiments on the NAS-bench-101 dataset (Ying et al., 2019)
Dataset Splits No The paper refers to using existing data for initialization (e.g., 'Tinit = 2000 evaluated network architectures'), but does not provide specific train/validation/test dataset splits in the context of model training for reproducibility.
Hardware Specification No run on a 16 core dual-CPU server
Software Dependencies No The experiments are implemented in python using the numpy, scikit-learn and botorch library
Experiment Setup Yes All algorithm use the hyper-parameters suggested by theory. When not applicable, cross validated parameters that perform the best for each individual algorithm are used (e.g. the kernel bandwidth). All the detailed choices and further experiments are reported in the Appendix D.