Zonotope Hit-and-run for Efficient Sampling from Projection DPPs

Authors: Guillaume Gautier, Rémi Bardenet, Michal Valko

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We investigate the behavior of our Algorithm 5 on synthetic graphs in Section 4.1, in summary extraction in Section 4.2, and on MNIST in Appendix A. Our empirical results demonstrate that this extends to sampling projection DPPs, i.e., our sampler is more sample-efficient than previous approaches which in turn translates to faster convergence when dealing with costly-to-evaluate functions, such as summary extraction in our experiments.
Researcher Affiliation Academia 1Univ. Lille, CNRS, Centrale Lille, UMR 9189 CRISt AL 2INRIA Lille Nord Europe, Seque L team.
Pseudocode Yes The paper includes multiple algorithm blocks, specifically: Algorithm 1 basis Exchange Sampler, Algorithm 2 unif Zono Hit And Run, Algorithm 3 extract Basis, Algorithm 4 vol Zono Hit And Run, and Algorithm 5 zonotope Sampler.
Open Source Code No The paper states that the implementation uses third-party tools (GLPK, CVXOPT) but does not provide any link or explicit statement about releasing its own source code for the methodology described.
Open Datasets Yes We investigate the behavior of our Algorithm 5 on synthetic graphs in Section 4.1, in summary extraction in Section 4.2, and on MNIST in Appendix A.
Dataset Splits No The paper does not provide specific details about training, validation, or test dataset splits (e.g., percentages or counts).
Hardware Specification No The paper mentions running experiments 'on a modern laptop' but does not provide specific hardware details such as CPU/GPU models, memory, or other specifications.
Software Dependencies No The paper mentions 'calling the simplex algorithm in the GLPK (Oki, 2012) solver with CVXOPT (Andersen et al., 2008) from Python'. However, it does not provide specific version numbers for GLPK, CVXOPT, or Python.
Experiment Setup Yes For Algorithm 5, a value of the linear objective c is drawn once and for all, for each graph, from a standard Gaussian distribution. We run both algorithms for 70 seconds, which corresponds to roughly 50 000 iterations of Algorithm 5. Moreover, we run 100 chains in parallel for each of the two algorithms. For each of the 100 repetitions, we initialize the two algorithms with the same random initial basis, obtained by solving (9) once, with x = Au and u U[0,1]n. We summarize this 64-sentence article as a subset of 11 sentences.