Skip to content

JonathanWenger/probabilistic-linear-solvers-for-ml

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

63 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Probabilistic Linear Solvers for Machine Learning

arXiv NeurIPS ProbNum ProbNum

This repository contains additional resources such as experiments and data for the paper Probabilistic Linear Solvers for Machine Learning by Jonathan Wenger and Philipp Hennig published at NeurIPS 2020.


Abstract: Linear systems are the bedrock of virtually all numerical computation. Machine learning poses specific challenges for the solution of such systems due to their scale, characteristic structure, stochasticity and the central role of uncertainty in the field. Unifying earlier work we propose a class of probabilistic linear solvers which jointly infer the matrix, its inverse and the solution from matrix-vector product observations. This class emerges from a fundamental set of desiderata which constrains the space of possible algorithms and recovers the method of conjugate gradients under certain conditions. We demonstrate how to incorporate prior spectral information in order to calibrate uncertainty and experimentally showcase the potential of such solvers for machine learning.


PLS Illustration

Implementation

An open-source implementation of our method is available as part of ProbNum.


probabilistic numericsProbNum implements probabilistic numerical methods in Python. Such methods solve numerical problems from linear algebra, optimization, quadrature and differential equations using probabilistic inference. This approach captures uncertainty arising from finite computational resources and stochastic input.


You can install ProbNum from the Python package index via:

pip install probnum==0.1.21

To get started check out the tutorials on how to use the probabilistic linear solver within ProbNum and the API documentation.

Experiments

You can reproduce all experiments and plots shown in the paper using ProbNum v0.1.21.

git clone git@github.com:JonathanWenger/probabilistic-linear-solvers-for-ml.git
cd probabilistic-linear-solvers-for-ml
pip install -r requirements.txt

Plots and Illustrations

Jupyter notebooks reproducing plots and illustrations are located in ./experiments/notebooks. Simply install Jupyter and run the notebooks.

jupyter notebook

Kernel Gram Matrix Inversion

Calibration experiments on kernel matrices generated from the flight delay dataset can be run in the following way.

cd experiments/scripts
python kernel_matrix_inversion.py

All calibration strategies reduced overconfidence of the solver for the evaluated kernel matrices.

calibration_experiment

Galerkin's Method

To apply the probabilistic linear solver to a discretization of a partial differential equation run the associated Jupyter notebook in ./experiments/notebooks.

If you wish to regenerate the mesh and resulting linear system via Galerkin's method yourself, begin by installing FeNiCS. On Ubuntu you can simply install from the package management system.

sudo apt-get install fenics

You can now generate the coarse and fine meshes and associated linear systems for the Dirichlet equation via:

cd experiments/scripts
python poisson_pde.py -r 6
python poisson_pde.py -r 128

Citation

If you use this work in your research, please cite our paper:

Jonathan Wenger and Philipp Hennig. Probabilistic Linear Solvers for Machine Learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020

@incollection{wenger2020problinsolve,
  author        = {Jonathan Wenger and Philipp Hennig},
  title         = {Probabilistic Linear Solvers for Machine Learning},
  booktitle 	= {Advances in Neural Information Processing Systems (NeurIPS)},
  year          = {2020},
  keywords      = {probabilistic numerics, numerical linear algebra, machine learning},
  url           = {https://github.com/JonathanWenger/probabilistic-linear-solvers-for-ml}
}

Acknowledgements

The authors gratefully acknowledge financial support by the European Research Council through ERC StG Action 757275 / PANAMA; the DFG Cluster of Excellence "Machine Learning - New Perspectives for Science", EXC 2064/1, project number 390727645; the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039A); and funds from the Ministry of Science, Research and Arts of the State of Baden-Württemberg.

JW is grateful to the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for support.