This repository contains the code for running and reproducing the experiments from the paper Sensitivity-Aware Amortized Bayesian Inference, published in Transactions on Machine Learning Research (arXiv version).
SA-ABI is a simulation-based method for fully amortized sensitivity analysis across all major dimensions of a Bayesian model: likelihood, prior, approximator, and data. This is achieved by training a deep ensemble of neural networks to amortize over a familiy of computational models.
The code depends on the BayesFlow library, which implements the neural network architectures and training utilities.
The article can be cited as:
@article{elsemueller2024sensitivity,
title={Sensitivity-Aware Amortized Bayesian Inference},
author={Lasse Elsem{\"u}ller and Hans Olischl{\"a}ger and Marvin Schmitt and Paul-Christian B{\"u}rkner and Ullrich Koethe and Stefan T. Radev},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2024},
url={https://openreview.net/forum?id=Kxtpa9rvM0},
}Contains the main analysis code for each experiment in self-contained Jupyter notebooks and supporting Python scripts:
- Experiment 1: COVID-19 Outbreak Dynamics
- Experiment 2: Climate Trajectory Forecasting
- Experiment 3: Hierarchical Models of Decision-Making
- Additional OOD Tests for Experiment 3
Contains custom Julia and Python functions that enable the analyses.
This work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under the research training group Statistical Modeling in Psychology (SMiP; GRK 2277) and under Germany’s Excellence Strategies EXC-2075 - 390740016 (the Stuttgart Cluster of Excellence SimTech) and EXC-2181 - 390900948 (the Heidelberg Cluster of Excellence STRUCTURES). Additionally, it was supported by the Google Cloud Research Credits program (award GCP19980904), the state of Baden-Württemberg through bwHPC, the Cyber Valley Research Fund (grant number: CyVy-RF-2021-16), and the Informatics for Life initiative funded by the Klaus Tschira Foundation.
MIT