Putting MRFs on a Tensor Train

Authors: Alexander Novikov, Anton Rodomanov, Anton Osokin, Dmitry Vetrov

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our methods on several datasets and compare them against several state-of-the-art techniques. We report a significant improvement in the problems of the estimation of the partition function and the marginal distributions and show comparable results in the MAP-inference problem. (Abstract) ... 7. Experiments In all our experiments we use a non-optimized MATLAB implementation3 of our methods.
Researcher Affiliation Academia 1 Moscow State University, Moscow, Russia 2 Higher School of Economics, Moscow, Russia
Pseudocode Yes Algorithm 1 Compute the partition function Z
Open Source Code Yes https://github.com/bihaqo/TT-MRF (footnote 3 on page 5)
Open Datasets Yes In our evaluation we mainly use the homogeneous and heterogeneous Ising models, and we refer to the supplementary material for the detailed description of the experimental setup. ... First we compare our method (TT) against the following methods implemented in the Lib DAI system (Mooij, 2010): ...
Dataset Splits No The paper describes generating models and evaluating them, but does not specify clear train/validation/test dataset splits (e.g., percentages, sample counts, or citations to predefined splits) needed to reproduce the data partitioning.
Hardware Specification No The paper states that experiments were run using a MATLAB implementation and provides running times, but no specific hardware details (such as CPU/GPU models or memory specifications) are mentioned.
Software Dependencies Yes In all our experiments we use a non-optimized MATLAB implementation3 of our methods. For operations related to the TT-format we use the TT-Toolbox4 implemented in MATLAB as well. ... 4http://spring.inm.ras.ru/osel/download/tt22.zip
Experiment Setup Yes For AIS method we select parameters (1000 intermediate distributions, 70 samples each) to maximize the accuracy achieved within 60 seconds per model. ... For each value of temperature T we generate 50 homogeneous 10x10 Ising models... Here pairwise weights are generated uniformly from [f, f] with parameter f varying from 0.25 to 3.