Comparing Amortized to MCMC-based Bayesian Inference for Cognitive Models of the Stop-Signal Paradigm

Amortized Bayesian Inference (ABI) is an emerging technique that improves on the ideas of approximate Bayesian computation by integrating non-parametric model learning via deep learning, only requiring a data-generative model. This makes it a promising approach for cognitive modelling where more complex models often lack algebraic solutions to enable standard Markov-chain Monte Carlo (MCMC) approaches to parameter estimation. However, while simulation studies show promising convergence for cognitive models with ABI, it is not clear which conditions are needed to ensure its usefulness. Furthermore, it is often not clear if marginal and joint posterior estimates are true reflections of Bayesian inference for complex statistical models. The presented research investigates how ABI compares to MCMC-based methods in the context of cognitive models of the stop-signal paradigm. Specifically, we investigated convergence and computational effort in ABI as implemented by BayesFlow (Radev et al., 2020) compared to the MCMC-based BEESTS as implemented by the DMC R package functions (Matzke et al., 2013; Heathcote et al., 2019). We present numerical comparisons for and draw conclusions and take-aways for the application of ABI for (ExGaussian) models for stop-signal detection tasks.