Efficient identification of informative features in simulation-based inference

Authors: Jonas Beck, Michael Deistler, Yves Bernaerts, Jakob H Macke, Philipp Berens

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For a simple linear Gaussian model and non-linear HH models, we show that the obtained posterior estimates are as accurate and robust as when repeatedly retraining density estimators from scratch, despite being much faster to compute. We then apply our algorithm to study which features are the most useful for constraining posteriors over parameters of the HH model.
Researcher Affiliation Academia Jonas Beck University of Tübingen jonas.beck@uni-tuebingen.de Michael Deistler University of Tübingen michael.deistler@uni-tuebingen.de Yves Bernaerts University of Tübingen yves.bernaerts@uni-tuebingen.de Jakob H. Macke University of Tübingen Max Planck Institute for Intelligent Systems jakob.macke@uni-tuebingen.de Philipp Berens University of Tübingen philipp.berens@uni-tuebingen.de
Pseudocode No The paper does not contain a pseudocode block or an explicitly labeled algorithm.
Open Source Code Yes We implement FSLM using python 3.8, building on top of the public sbi library [36]. All the code is available at github.com/berenslab/fslm_repo.
Open Datasets No The paper uses simulated data generated from a mechanistic model rather than a publicly available dataset. "Parameters are sampled from the prior and used to simulate a synthetic dataset." and "To train the density estimator we simulated 1 million observations with parameters drawn from a Uniform prior over biologically plausible parameter ranges (see A.3)."
Dataset Splits Yes Training was terminated when the log-likelihoods had stopped improving on a validation set that was split from the training set at 10% over 20 epochs.
Hardware Specification Yes All computations were done on an internal cluster running Intel(R) Xeon(R) Gold 6226R CPUs @ 2.90GHz.
Software Dependencies No The paper mentions 'python 3.8' and 'sbi library [36]' but only provides a version number for Python, not for other key software components or libraries.
Experiment Setup Yes The MDNs had three hidden layers and 10 mixture components. The MAF consisted of five 3 layer deep MADEs [43]. Training was done on 10,000 samples, then 500 samples were drawn from each posterior, for a total of 10 different initializations. Training was terminated when the log-likelihoods had stopped improving on a validation set that was split from the training set at 10% over 20 epochs.