On Fast Leverage Score Sampling and Optimal Learning

Authors: Alessandro Rudi, Daniele Calandriello, Luigi Carratino, Lorenzo Rosasco

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments Leverage scores accuracy. We first study the accuracy of the leverage scores generated by BLESS and BLESS-R, comparing SQUEAK [8] and Recursive-RLS (RRLS) [9]. We begin by uniformly sampling a subsets of n = 7 104 points from the SUSY dataset [19], and computing the exact leverage scores (i, λ) using a Gaussian Kernel with σ = 4 and λ = 10 5, which is at the limit of our computational feasibility. We then run each algorithm to compute the approximate leverage scores e JH(i, λ), and we measure the accuracy of each method using the ratio e JH(i, λ)/ (i, λ) (R-ACC). The final results are presented in Figure 1.
Researcher Affiliation Academia Alessandro Rudi INRIA Sierra team, Daniele Calandriello LCSL IIT & MIT, Genoa, Italy Luigi Carratino University of Genoa, Genoa, Italy Lorenzo Rosasco University of Genoa, LCSL IIT & MIT
Pseudocode Yes Algorithm 1 Bottom-up Leverage Scores Sampling (BLESS) and Algorithm 2 Bottom-up Leverage Scores Sampling without Replacement (BLESS-R)
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a code repository for the described methodology.
Open Datasets Yes We begin by uniformly sampling a subsets of n = 7 104 points from the SUSY dataset [19]
Dataset Splits No The paper uses the SUSY and HIGGS datasets but does not explicitly provide specific training/validation/test dataset split information such as percentages, sample counts, or a detailed splitting methodology.
Hardware Specification Yes We gratefully acknowledge the support of NVIDIA Corporation for the donation of the Titan Xp GPUs and the Tesla k40 GPU used for this research.
Software Dependencies No The paper does not provide specific software dependency details, such as library names with version numbers (e.g., 'PyTorch 1.9', 'Python 3.8').
Experiment Setup Yes For the SUSY dataset we use a Gaussian Kernel with σ = 4, λfalkon = 10 6, λbless = 10 4 obtaining MH ' 104 Nyström centres. For the HIGGS dataset we use a Gaussian Kernel with σ = 22, λfalkon = 10 8, λbless = 10 6, obtaining MH ' 3 104 Nyström centres.