Fast ε-free Inference of Simulation Models with Bayesian Conditional Density Estimation

Authors: George Papamakarios, Iain Murray

NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We showcase three versions of our approach: (a) learning the posterior with Algorithm 2 where qφ is a conventional MDN and the proposal prior p(θ) is taken to be the actual prior p(θ), which we refer to as MDN with prior; (b) training a proposal prior with Algorithm 1 where qφ is an MDN-SVI, which we refer to as proposal prior; and (c) learning the posterior with Algorithm 2 where qφ is an MDN-SVI and the proposal prior p(θ) is taken to be the one learnt in (b), which we refer to as MDN with proposal. We compare to three ABC baselines: (a) rejection ABC [21], where parameters are proposed from the prior and are accepted if x xo < ϵ; (b) MCMC-ABC [13] with a spherical Gaussian proposal, whose variance we manually tuned separately in each case for best performance; and (c) SMC-ABC [2, 5]...
Researcher Affiliation Academia George Papamakarios School of Informatics University of Edinburgh g.papamakarios@ed.ac.uk Iain Murray School of Informatics University of Edinburgh i.murray@ed.ac.uk
Pseudocode Yes Algorithm 1: Training of proposal prior
Open Source Code Yes Code for reproducing the experiments is provided in the supplementary material and at https://github.com/gpapamak/epsilon_free_inference.
Open Datasets No The paper uses simulated data generated by the models themselves (e.g., 'randomly generated observations xo from the model' for Bayesian linear regression, and details in supplementary material for Lotka Volterra and M/G/1), rather than referencing a publicly available dataset with concrete access information.
Dataset Splits No The paper discusses the concept of a 'validation set' in the context of avoiding overfitting, but does not specify actual train/validation/test splits used for its experiments (e.g., percentages or sample counts).
Hardware Specification No The paper does not specify any particular hardware components (e.g., GPU models, CPU types, or memory) used for running the experiments.
Software Dependencies No The paper mentions 'Adam [11]' as the optimizer used for training MDNs but does not provide specific version numbers for Adam or any other key software components or libraries.
Experiment Setup Yes All MDNs were trained using Adam [11] with its default parameters. and All MDNs have one hidden layer with 20 tanh units and 2 Gaussian components, except for the proposal prior MDN which has a single component.