Tensor Belief Propagation

Authors: Andrew Wrigley, Wee Sun Lee, Nan Ye

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare our algorithm experimentally with several approximate inference algorithms and show that it performs well. We present experiments on grid-structured Ising models, random graphs with pairwise Ising potentials, and two real-world datasets from the UAI 2014 Inference Competition (Gogate, 2014).
Researcher Affiliation Academia 1Australian National University, Canberra, Australia. 2National University of Singapore, Singapore. 3Queensland University of Technology, Brisbane, Australia.
Pseudocode Yes Algorithm 1 Tensor Belief Propagation
Open Source Code No The paper states that TBP was implemented in C++ inside the lib DAI framework, but does not provide a link or explicit statement that their specific implementation of TBP is open source.
Open Datasets Yes We present experiments on grid-structured Ising models, random graphs with pairwise Ising potentials, and two real-world datasets from the UAI 2014 Inference Competition (Gogate, 2014). UAI 2014 Inference Competition. http://www.hlt.utdallas.edu/ vgogate/ uai14-competition/index.html, 2014.
Dataset Splits No The paper does not explicitly specify training, validation, and test dataset splits for the experiments. It describes how models are generated or uses pre-defined competition problem instances rather than splitting a dataset.
Hardware Specification Yes all tests were executed on a single core of a 1.4 GHz Intel Core i5 processor.
Software Dependencies Yes TBP was implemented in C++ inside the lib DAI framework using Eigen (Guennebaud et al., 2010). lib DAI: A Free and Open Source C++ Library for Discrete Approximate Inference in Graphical Models. Journal of Machine Learning Research, 11: 2169–2173, August 2010. Eigen v3. http://eigen.tuxfamily.org, 2010.
Experiment Setup Yes In our experiments, we choose the wij uniformly from [-2, 2] (mixed interactions) or [0, 2] (attractive interactions), and the bi uniformly from [-1, 1]. We use a symmetric rank-2 tensor decomposition for the pairwise potentials. Parameters used for BP, MF, TRW and Gibbs are given in the supplementary material. The initial potential functions are decomposed into mixtures with r components. We show results for r = 2 and r = 4.