Deep Learning for Computational Cognitive Modeling

Amortized Bayesian inference (ABI) with neural networks has accelerated and transformed cognitive modeling workflows. However, the utility of ABI for rapid iteration over varying modeling assumptions remains limited: changing parameterizations, generative functions, priors, and design variables all necessitate model retraining and hence diminish the benefits of amortization. To address these issues, we propose the first meta-amortized, multi-task learning framework for cognitive modeling. Our framework induces amortized neural estimators that remain valid for entire model classes under a generalized linear model (GLM), allowing for varying numbers of regressors, model parameters, and sample sizes as well as test-time adaptation of modeling assumptions. Our framework is bootstrapped by a novel transformer-based architecture that enables fully Bayesian inference without limiting parametric assumptions on the resulting posterior. We present promising quantitative results on families of decision-making models for binary, multi-alternative, and continuous responses, showing a minimal amortization gap at the price of a vastly expanded generalization scope. This work presents the first step towards an interpretable foundation model for Bayesian cognitive modeling.

This is an in-person presentation on July 19, 2026 (09:00 ~ 09:20 EDT).

Cognition is inherently dynamic. Cognitive states fluctuate over time due to learning, fatigue, strategic adaptation, and other influences. Yet most mechanistic cognitive models assume time-invariant or stationary parameters, effectively ignoring non-stationary changes. Our recently introduced superstatistics framework allows cognitive model parameters to evolve over time using flexible transition models. This flexibility, however, introduces new challenges in specifying, fitting, and validating models. To facilitate broader adoption, we developed superstats, a flexible and user-friendly software package that allows modelers to easily apply the superstatistics framework to any cognitive model. The package provides an interface for selecting and combining predefined transition models, supporting gradual shifts or sudden jumps. Furthermore, it enables fast estimation and prediction of parameter trajectories via amortized Bayesian inference and includes diagnostic tools such as parameter recovery and simulation-based calibration. We demonstrate its application using the diffusion decision model and show that key parameters often exhibit substantial and systematic temporal variation.

This is an in-person presentation on July 19, 2026 (09:20 ~ 09:40 EDT).

Race models formalize and explain individual differences in speeded decision making. They assign a runner to each competing response option that tracks accumulated evidence in its favor. Race model parameters can be estimated through the model likelihood that combines the probability and cumulative density functions of each runner. However, Bayesian inference for complex race models, whose runners have intractable densities, with traditional numerical methods has been challenging. We propose a flexible neural estimation method that decouples training from the choice of prior distribution or model parameterization. Instead of training a complex neural estimator for the full model posterior, we combine a small neural network with a monotonic flow to efficiently estimate the conditional probability density and cumulative density functions. We then combine both functions to estimate race model likelihoods that can be used in tandem with modern MCMC samplers. We validate our approach and demonstrate its computational feasibility using two exemplary race models: The Racing Diffusion Model (RDM) and a conflict task version of the RDM with intractable runner densities. We show that our method can estimate both single-subject and hierarchical models, highlighting that neural density estimation in tandem with MCMC can be a simple yet powerful alternative to full-posterior neural estimation. Based on our results, we argue that, when ad hoc flexibility for choosing prior distributions or model parameterizations is needed, learning the model likelihood can have advantages over end-to-end amortized Bayesian inference.

This is an in-person presentation on July 19, 2026 (09:40 ~ 10:00 EDT).

Human decision-making is widely understood as a dynamic process of evaluating information and accumulating evidence over time. Represented by sequential sampling models, theory-driven computational approaches offer well-specified accounts of the dynamics that generate choice behaviors. However, because such models must be specified a priori, they typically explore a restricted set of hypothesized mechanisms and may miss structure not anticipated by the modeler. This has motivated growing interest in data-driven approaches that can complement theory-driven modeling by discovering latent structure directly from behavioral data. The present work develops an unsupervised learning framework to infer latent factors and generative structure directly from choice and response time data. The approach adopts a Neural Process architecture in which an encoder performs amortized variational inference over participant-level data to estimate latent factors, and a decoder then conditions on these participant-level latent variables together with trial-wise cues to reconstruct both choices and response times. In Simulation Study 1, we showed that this approach reconstructed choices and response times and recovered both the true dimensionality of the parameter space and the underlying parameter values. Simulation Study 2 extended these results to settings with trial-wise cues. In an application to real data, we applied the model to a large-scale, multi-task study of risk preference and identified latent risk-related factors that generalize across tasks. Together, these results introduce a data-driven framework for jointly modeling choices and response times, enabling the identification of latent structure in decision-making data beyond what is prespecified by traditional cognitive models.

This is an in-person presentation on July 19, 2026 (10:00 ~ 10:20 EDT).