Relaxations for inference in restricted Boltzmann machines

Authors: Sida I. Wang; Roy Frostig; Percy Liang; Christopher D. Manning

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experiment on MAP inference tasks in several restricted Boltzmann machines. We also use our underlying sampler to estimate the log-partition function of restricted Boltzmann machines and compare against other sampling-based methods.
Researcher Affiliation Academia Sida Wang sidaw@cs.stanford.edu Roy Frostig rf@cs.stanford.edu Percy Liang pliang@cs.stanford.edu Christopher D. Manning manning@stanford.edu Computer Science Department, Stanford University, Stanford, CA 94305, USA
Pseudocode Yes Algorithm 1: Randomized relax-and-round MAP sampler (rrr-MAP).
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing the code for the work described, nor does it provide a direct link to a source-code repository.
Open Datasets Yes MNIST. We downloaded the weights for an RBM over the {0, 1}n domain, trained by Salakhutdinov & Murray (2008) to model the MNIST dataset distribution.
Dataset Splits No The paper does not provide specific details about training, validation, or test dataset splits (e.g., percentages, sample counts, or defined split methodologies). It mentions using pre-trained weights for MNIST and randomly generated parameters for other RBMs, implying no explicit data splitting was performed for their specific experiments.
Hardware Specification No The paper mentions that 'solving a semidefinite program of such size takes hours on modern hardware' but does not provide specific details about the exact GPU/CPU models, processor types, or detailed computer specifications used for running its experiments.
Software Dependencies No The paper mentions using an 'off-the-shelf IQP solver (Gurobi)' but does not provide specific version numbers for Gurobi or any other ancillary software components.
Experiment Setup Yes Executions of rrr-MAP use LRP2 as the initial relaxation (i.e. width k = 2). Gurobi is given an execution time limit that is 10x that of rrr. AG is an annealed Gibbs procedure with a linear temperature schedule.