Skip to content

Contains the code for reproducing the experiments and results of the paper "Neural Superstatistics: A Bayesian Method for Estimating Dynamic Models of Cognition".

License

Notifications You must be signed in to change notification settings

bayesflow-org/Neural-Superstatistics

Repository files navigation

Neural Superstatistics

This repository contains the code for running the experiments and reproducing all results reported in our paper Neural Superstatistics for Bayesian Estimation of Dynamic Cognitive Models. We employ the superstatistics perspective to augment mechanistic cognitive models with a temporal dimension and perform amortized estimation of the resulting dynamics.

The details of the method are described in our paper:

Schumacher, L., Bürkner, P. C., Voss, A., Köthe, U., & Radev, S. T. (2022). Neural Superstatistics for Bayesian Estimation of Dynamic Cognitive Models arXiv preprint arXiv:2211.13165, available for free at: https://arxiv.org/abs/2211.13165.

The code depends on the BayesFlow library, which implements the neural network architectures and training utilities.

Cite

@article{schumacher2022neural,
      title={Neural Superstatistics: A Bayesian Method for Estimating Dynamic Models of Cognition}, 
      author={Lukas Schumacher and Paul-Christian Bürkner and Andreas Voss and Ullrich Köthe and Stefan T. Radev},
      year={2022},
      journal={arXiv preprint arXiv:2211.13165}
}

Estimating time-varying parameters

The following animation illustrates parameter estimation of a drift-diffusion model (DDM) with time-varying parameters over 3200 time steps:

All applications are structured as self-contained Jupyter notebooks, which are detailed below.

Benchmark studies:

  • Bayesloop benchmark: Comparison of our neural estimation method with the Bayesloop method which specializes on estimating time-varying parameters in low-dimensional problems.
  • Stan benchmark: Comparison of our neural estimation method with HMC-MCMC as implemented in Stan.

Simulation studies:

  • Simulation study: Assessment of the parameter recovery performance of a non-stationary drift-diffusion model (DDM) in four different simulation scenarios.

Human data applications:

  • Optimal policy: Fitting a non-stationary DDM to data from a random-dot motion task.
  • Lexical decision: Fitting a non-stationary DDM to long human time series data from a lexical decision task.

Support

This work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation; grant number GRK 2277 ”Statistical Modeling in Psychology”), by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC-2075 - 390740016 (the Stuttgart Cluster of Excellence SimTech), and by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC-2181 - 390900948 (the Heidelberg Cluster of Excellence STRUCTURES).

License

MIT

About

Contains the code for reproducing the experiments and results of the paper "Neural Superstatistics: A Bayesian Method for Estimating Dynamic Models of Cognition".

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published