Bayesian Active Learning with Fully Bayesian Gaussian Processes

Authors: Christoffer Riis, Francisco Antunes, Frederik Hüttel, Carlos Lima Azevedo, Francisco Pereira

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

Reproducibility Variable Result LLM Response
Research Type Experimental Across six simulators, we empirically show that B-QBC, on average, achieves the best marginal likelihood, whereas QB-MGP achieves the best predictive performance. We show that incorporating the bias-variance trade-off in the acquisition functions mitigates unnecessary and expensive data labeling.
Researcher Affiliation Academia Christoffer Riis Francisco Antunes Frederik Boe Hüttel Carlos Lima Azevedo Francisco Câmara Pereira DTU Management Machine Learning for Smart Mobility {chrrii,franant,fbohy,climaz,camara}@dtu.dk
Pseudocode Yes See Appendix A.6 for the formulas and the pseudo-code.
Open Source Code Yes The code for reproducing the experiments is available on Git Hub.2 https://github.com/coriis/active-learning-fbgp
Open Datasets Yes All of them can be found at https://www.sfu.ca/~ssurjano/ [Surjanovic and Bingham, 2022].
Dataset Splits No The paper describes data acquisition for active learning (e.g., initial data sets, unlabeled pool) but does not specify a conventional train/validation/test dataset split.
Hardware Specification Yes With seven simulators and five acquisition functions, this gives 350 active learning runs, each with a running time on approximately one hour, using five CPU cores on a Threadripper 3960X.
Software Dependencies No The paper mentions software like NUTS, Pyro, and GPy Torch, but does not provide specific version numbers for these dependencies.
Experiment Setup Yes In each iteration of the active learning loop, the inputs are rescaled to the unit cube [0, 1]d, and the outputs are standardized to have zero mean and unit variance. Following Lalchand and Rasmussen [2020], we give all the hyperparameters relatively uninformative N(0, 3) priors in log space. The initial data sets consist of three data points chosen by maximin Latin Hypercube Sampling [Pedregosa et al., 2011], and in each iteration, one data point is queried. ... The inference in FBGP is carried out using NUTS [Hoffman and Gelman, 2014] in Pyro [Bingham et al., 2019] with five chains and 500 samples, including a warm-up period with 200 samples. The remaining 1500 samples are all used for the acquisition functions.