Researchers crack speed bottleneck in Bayesian inference

A team of researchers from KU Leuven and University of California has developed a novel divide-and-conquer framework that dramatically improves the scalability of amortized Bayesian inference (ABI) for drift diffusion models (DDM), a cornerstone in cognitive decision-making research. The method exploits the DDM's independence assumption by decomposing full datasets into pairwise shards, each processed by a shared neural network. Posteriors from these shards are then merged using consensus MCMC to approximate the full posterior distribution.

In benchmarks using simulated datasets, the approach delivered accuracy and uncertainty quantification comparable to traditional Markov Chain Monte Carlo (MCMC) methods while reducing computational costs by several orders of magnitude. This breakthrough not only addresses the long-standing challenge of generalizing ABI across diverse study designs but also offers a blueprint for scalable inference in other complex stochastic models beyond cognitive science applications.