HEALING PRODUCTS OF GAUSSIAN PROCESS EXPERTS - ICML 2020

In this repository can be found implementation of the aggregations and weighting approaches for products of Gaussian process experts proposed in our paper [1], along with several previously proposed PoEs.

Aggregation Methods:

• PoE
• gPoE
• BCM
• rBCM
• Barycenter

Along with other baselines:

• Full GP
• Linear regression

Weighting Methods

• Differential Entropy
• Softmax-Variance
• Uniform
• No-weights

How to Run Experiments:

Unzip the airline dataset Code/bayesian_benchmarks_modular/bayesian_benchmarks/data/airline/DelayedFlights_all.csv.zip

Please move into Code/bayesian_benchmarks_modular and run:

• python -m pytest bayesian_benchmarks/scripts/run_all_pytest.py -n X

where X is the number of experiments ran in parallel.

Results can be viewed in Code/bayesian_benchmarks_modular/bayesian_benchmarks/results/view_results.ipynb

Dependencies:

• Tensorflow 2.0
• GPflow 2.0.1
• Numpy 0.18.1
• Pandas 0.25.1
• pytest-xdist
• tqdm 4.32.1
• sklearn 0.21.2

Our Paper

This repository complements our paper:

[1] Healing Products of Gaussian Process Experts, Samuel Cohen, Rendani Mbuvha, Tshilidzi Marwala, Marc Deisenroth, International Conference in Machine Learning 2020