Large-scale log-determinant computation through stochastic Chebyshev expansions

Authors: Insu Han, Dmitry Malioutov, Jinwoo Shin

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments
Researcher Affiliation Collaboration Insu Han HAWKI17@KAIST.AC.KR Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, Korea Dmitry Malioutov DMALIOUTOV@US.IBM.COM Business Analytics and Mathematical Sciences, IBM Research, Yorktown Heights, NY, USA Jinwoo Shin JINWOOS@KAIST.AC.KR Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, Korea
Pseudocode Yes Algorithm 1 Log-determinant approximation for positive definite matrices with σmax < 1
Open Source Code Yes Our code is at http://sites.google.com/site/mijirim/logdet_code.zip
Open Datasets Yes We use the data-set from (Aune et al., 2014) that provides satellite measurements of ozone levels over the entire earth following the satellite tracks.
Dataset Splits No The paper discusses using data for training (e.g., 'estimate the log-likelihood as a function of ρ'), but does not provide specific training/validation/test dataset splits (percentages, counts, or citations to predefined splits).
Hardware Specification Yes We use a machine with 3.40 Ghz Intel I7 processor with 24 GB RAM.
Software Dependencies No The paper mentions using specific algorithms and code (e.g., 'QUIC algorithm (Hsieh et al., 2013)') and provides its own code link, but it does not specify version numbers for any key software components or libraries.
Experiment Setup Yes We generate a random matrix C Rd d, where the number of non-zero entries per each row is around 10. We first select five non-zero off-diagonal entries in each row with values uniformly distributed in [ 1, 1]. To make the matrix symmetric, we set the entries in transposed positions to the same values. Finally, to guarantee positive definiteness, we set its diagonal entries to absolute row-sums and add a small weight, 10 3. ... where we choose m = 10, n = 15, σmin = 10 3 and σmax = C 1.