Amortized Bayesian Workflow of the Diffusion Model for Conflict Tasks

This repository contains the work of Schaefer et al. (2025).

We propose a generalized Amortized Bayesian Workflow to yield optimal performance in the parameter estimation of cognitive models as well as pretrained networks for Amortized Bayesian Inference of the Diffusion Model for Conflict Tasks (Ulrich et al., 2015).

🔧 Project Status

This project is currently under construction.

🚀 Getting Started

Download the Code

First, open a terminal on your local machine and navigate to the directory where you want to store the project. Then, run the following command to clone the repository to your local machine:

git clone https://github.com/simschaefer/amortized_dmc.git

Alternatively, download the repository by clicking at Code on the right hand side, click Download ZIP and extract the .zip file in the directory where you want to store the project.

Install Dependencies

All packages that were used in this project are listed in requirements.txt. Create a fresh conda environment amortized_dmc:

conda create --name amortized_dmc python=3.11.11

and activate the environment:

conda activate amortized_dmc

Install pip:

conda install pip

And all dependencies:

pip install -r requirements.txt

Tutorial Notebooks

To guide you through the key steps of our analyses, we provide comprehensive Jupyter notebooks covering the following topics:

• DMC data simulation
• Hyperparameter optimization
• Application of pretrained networks on empirical data

📁 Repository Structure

• dmc/

  Includes the simulator function DMC() in dmc_simulator.py as well as helper functions to fit empirical data in dmc_helpers.py

• empirical_data/

  Includes data from the experiment seperately for each spacing condition:

  • Narrow spacing: experiment_data_narrow.csv
  • Wide spacing: experiment_data_wide.csv

• model_specs/

  Stored information about training hyperparameters, network hyperparameters and simulator specifications used for the training of all networks. The names correspond with those of the pretrained networks.

• notebooks/

  Comprehensive examples for

  • Data simulation using the DMC simulator
  • Running automated hyperparameter optimization using optuna
  • Apply pretrained networks to empirical data

• optuna_results/

  Results from the Hyperparameter Optimization Phase using scripts/dmc_optuna.py.

• training/

  Includes the scripts used in the Training Phases to train all four networks based on either initial or updated priors and either including or excluding trial-to-trial variability of the non-decision time.

• training_checkpoints/

  Includes the checkpoints of the trained networks for each condition.

• plot_scripts/

  All scripts that were used to create the plots in the paper:

  • prior_predictive_check.py: Prior Predictive Checks of initial and updated priors against empirical data.

  • In Silico Evaluation Phase:

    • diagnostics.py: Computation of Recovery, Simulation-Based Calibration and Posterior Contraction for a fixed number of trials.
    • metrics_num_obs.py: Computation of all in silico metrics for a varying number of trials between 50 and 1000.
    • plots_metrics_num_obs.ipynb: Plotting data computed by metrics_num_obs.py

  • Application to Empirical Data Phase:

    • experimental_effects.py: Computation of standardized mean differences between experimental conditions (narrow vs. wide stimuli spacing)

    • posterior_predictive_check.py: Posterior Predictive Checks of individual RT and Accuracy Data, as well as correlation between mean RT, RT quantiles and mean accuracy.

    • posterior_predictive_check_delta_functions.py: Posterior Predictive Check of individual delta functions.

    • posterior_reliability.py: Split-Half correlation between individual parameter estimates for seven data sets.

    • reliability_comparison_plots.ipynb: Plotting data from posterior_reliability.py.

• scripts/

  Additional scripts:

  • dmc_optuna.py: automated hyperparameter optimization
  • prior_updating.py: updating of priors in the Prior Updating Phase using the networks trained on initial priors
  • simulate_data.py: data simulation used in the Benchmarking Phase
  • drift_dm_fitting_mogon.R: Parameter estimation for simulated data (simulate_data.py) using dRiftDM