unifying orthogonal monte carlo methods
play

Unifying Orthogonal Monte Carlo Methods From Kacs Random Walks To - PowerPoint PPT Presentation

Unifying Orthogonal Monte Carlo Methods From Kacs Random Walks To Hadamard Multi Rademachers Krzysztof Choromanski, Mark Rowland Wenyu Chen, Adrian Weller The Phenomenon of Orthogonal Monte Carlo Estimators Estimation task: Applications:


  1. Unifying Orthogonal Monte Carlo Methods From Kac’s Random Walks To Hadamard Multi Rademachers Krzysztof Choromanski, Mark Rowland Wenyu Chen, Adrian Weller

  2. The Phenomenon of Orthogonal Monte Carlo Estimators Estimation task: Applications: ● dimensionality reduction (JLT-mechanisms) isotropic distribution ● scaling kernel methods (e.g. Gaussian) (random feature maps) ● hashing algorithms (e.g. LSH) ● (sliced) Wasserstein Standard MC approach: distances (WGANs, autoencoders...) ● reinforcement learning (ES algorithms) ● and many, many more...

  3. The Phenomenon of Orthogonal Monte Carlo Estimators Sampling from the Estimation task: Haar measure on the O(d) group isotropic distribution (e.g. Gaussian) Expensive: O(n^3 time) The Orthogonal Trick: guarantees unbiasedness # of samples of the often implies better MC estimator <= dim accuracy

  4. Towards Computational Efficiency: The Zoo of Approximate MCs

  5. Towards Computational Efficiency: The Zoo of Approximate MCs ...

  6. Towards Computational Efficiency: The Zoo of Approximate MCs ...

  7. Towards Computational Efficiency: The Zoo of Approximate MCs ... ...

  8. Towards Computational Efficiency: The Zoo of Approximate MCs ... ...

  9. Towards Computational Efficiency: The Zoo of Approximate MCs ... ...

  10. Towards Computational Efficiency: The Zoo of Approximate MCs ... size size N x N N/2 x N/2 ... Constraints: ● ●

  11. Towards Computational Efficiency: The Zoo of Approximate MCs ... size size N x N N/2 x N/2 ... Constraints: ● ●

  12. Towards Computational Efficiency: The Zoo of Approximate MCs ... size size N x N N/2 x N/2 ... Constraints: ● ●

  13. On the Hunt for the Unifying Theory: The World of Givens Reflections and Rotations Kac’s random walk matrices Givens rotations Hadamard-Rademacher Chains Givens reflections reflection in the jth coordinate

  14. On the Hunt for the Unifying Theory: The World of Givens Reflections and Rotations Kac’s random walk matrices Hadamard-Rademacher Chains

  15. On the Hunt for the Unifying Theory: The World of Givens Reflections and Rotations Hadamard-MultiRademachers Butterfly Matrices

  16. First Theoretical Results for Free-Lunch Phenomenon in the Nonlinear Regime

  17. First Theoretical Results for Free-Lunch Phenomenon in the Nonlinear Regime Still more accurate estimator than unstructured MC baseline

  18. First Theoretical Results for Free-Lunch Phenomenon in the Nonlinear Regime Log-Linear Time Complexity (unstructured MC baseline has quadratic)

  19. First Theoretical Results for Free-Lunch Phenomenon in the Nonlinear Regime Analysis of the Total Variation Distance between Haar measure on d-sphere and measure induced by standard Kac’s random walk on d-sphere estimator estimated value Pillai, Smith 2016 Kac’s random walk on d-sphere mixes in O(d log d) steps

  20. First Theoretical Results for Free-Lunch Phenomenon in the Nonlinear Regime Analysis of the Total Variation Distance between Haar measure on d-sphere and measure induced by standard Kac’s random walk on d-sphere estimator estimated value Pillai, Smith 2016 Kac’s random walk on d-sphere mixes in O(d log d) steps More careful analysis of the LHS

  21. Maximum Mean Discrepancy Experiment How Does It Work In Practice ? Kernel Approximation via Random Features Reinforcement Learning via ES-methods Accuracy Computational Efficiency

  22. Maximum Mean Discrepancy Experiment How Does It Work In Practice ? Kernel Approximation via Random Features Reinforcement Learning via ES-methods Accuracy Computational Efficiency

  23. Thank you for your attention !

Recommend


More recommend