In cognitive neuroscience, computational modeling can formally adjudicate between theories and affords quantitative fits to behavioral/brain data. Candidate models are often defined as generative: models from which we can simulate data. However, typically cognitive neuroscience studies require going the other way around, from data to inferring models and their parameters. Common software tools for doing this inference typically depend on having access to a likelihood function which returns the likelihood of data observation for a model parameterization. In practice, the space of plausible generative models considered is dramatically constrained to the set of models with known likelihood functions, otherwise inference is often quite simply too costly to carry out. As a result, standard models are evaluated for convenience, even when other models might be better suited to describing how the brain generates cognition and behavior. Researchers at Brown developed a novel method which uses neural networks to learn approximate likelihoods for arbitrary generative models, allowing fast bayesian inference with only a one-off cost for model simulations that is amortized for future inference. They show that by relying on likelihood approximation networks bayesian inference can be accurately carried out for a variety of neurocognitive process models. The researchers provide code, a toolbox and a tutorial, allowing users to easily deploy these methods for arbitrary models linking the brain to behavioral data.
The original Python toolbox for hierarchical Bayesian parameter estimation of the Drift Diffusion Model, which was developed in Brown's Laboratory of Neural Computation and Cognition to link brain/behavioral data in health and disease, is widely used (more than 400 publications have used it and the userbase contributes to a very active listserv). However, that toolbox focused on a single model — the drift diffusion model of decision making. The researchers' novel tool can be deployed for arbitrary models in a user-friendly but rigorous way. Thus, it promises to substantially enhance discovery of more realistic brain dynamics underlying a variety of cognitive and behavioral phenomena, as well as how they are altered in disorders of brain function.