HNPE: Leveraging Global Parameters for Neural Posterior Estimation

Authors: Pedro Rodrigues, Thomas Moreau, Gilles Louppe, Alexandre Gramfort

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate quantitatively our proposal on a motivating example amenable to analytical solutions and then apply it to invert a well known non-linear model from computational neuroscience. All experiments described next are implemented with Python
Researcher Affiliation Academia Pedro L. C. Rodrigues Inria, CEA, Université Paris-Saclay, France Thomas Moreau Inria, CEA, Université Paris-Saclay, France Gilles Louppe University of Liège, Belgium Alexandre Gramfort Inria, CEA, Université Paris-Saclay, France
Pseudocode Yes Algorithm 1: Sequential posterior estimation for hierarchical models with global parameters
Open Source Code Yes Code is available in the supplementary materials. The code required for reproducing most of the results presented in the paper is available at https://github.com/plcrodrigues/HNPE
Open Datasets Yes Data consists of recordings taken from a public dataset (Cattan et al., 2018)
Dataset Splits No The paper does not explicitly provide specific training/validation/test dataset splits with percentages or sample counts for reproducibility.
Hardware Specification No This work was granted access to the HPC resources of IDRIS under allocations 2021-AD011011172R1 made by GENCI.
Software Dependencies No All experiments described next are implemented with Python (Python Software Fundation, 2017) and the sbi package (Tejero-Cantero et al., 2020) combined with Py Torch (Paszke et al., 2019), Pyro (Bingham et al., 2018) and nflows (Durkan et al., 2020a) for posterior estimation
Experiment Setup Yes In all experiments, we use the Adam optimizer (Kingma and Ba, 2014) with default parameters, a learning rate of 5.10 4 and a batch size of 100. Our approximation to the posterior distribution consists of two conditional neural spline flows of linear order (Durkan et al., 2019), qφ1 and qφ2, both conditioned by dense neural networks with one layer and 20 hidden units. The normalizing flows qφ1 and qφ2 used in our approximations are masked autoregressive flows (MAF) (Papamakarios et al., 2017) consisting of three stacked masked autoencoders (MADE) (Germain et al., 2015), each with two hidden layers of 50 units, and a standard normal base distribution as input to the normalizing flow.