Group equivariant neural posterior estimation
Authors: Maximilian Dax, Stephen R Green, Jonathan Gair, Michael Deistler, Bernhard Schölkopf, Jakob H. Macke
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | As a real-world application, we use GNPE for amortized inference of astrophysical binary black hole systems from gravitational-wave observations. We show that GNPE achieves state-of-the-art accuracy while reducing inference times by three orders of magnitude. |
| Researcher Affiliation | Academia | Maximilian Dax Max Planck Institute for Intelligent Systems Tübingen, Germany maximilian.dax@tuebingen.mpg.de Stephen R. Green Max Planck Institute for Gravitational Physics Potsdam, Germany stephen.green@aei.mpg.de Jonathan Gair Max Planck Institute for Gravitational Physics Potsdam, Germany Michael Deistler Machine Learning in Science, University of Tübingen Tübingen, Germany Bernhard Schölkopf Max Planck Institute for Intelligent Systems Tübingen, Germany Jakob H. Macke Max Planck Institute for Intelligent Systems & Machine Learning in Science, University of Tübingen Tübingen, Germany |
| Pseudocode | No | The paper describes the algorithm in text and uses a diagram to illustrate the process, but does not include a formal pseudocode or algorithm block. |
| Open Source Code | Yes | We also provide the code at https://tinyurl.com/wmbjajv8. |
| Open Datasets | No | We train the inference network with a data set of 5 106 waveforms with parameters θ sampled from the priors specified in table D.1. The training data is generated via simulation as described in the paper, not taken from an existing public dataset for which access information is provided. |
| Dataset Splits | Yes | We train the inference network with a data set of 5 106 waveforms with parameters θ sampled from the priors specified in table D.1, and reserve 2% of the data for validation. |
| Hardware Specification | Yes | With batch size 4,096, training takes 16-18 days on a NVIDIA Tesla V100 GPU. |
| Software Dependencies | No | The paper lists software used (Py Torch, nflows, sbi, matplotlib, Chain Consumer) but does not provide specific version numbers for these software packages. |
| Experiment Setup | Yes | We pretrain the network with learning rate of 3 10 4 for 300 epochs with fixed PSD, and finetune for another 150 epochs with learning rate of 3 10 5 with varying PSDs. With batch size 4,096, training takes 16-18 days on a NVIDIA Tesla V100 GPU. |