Deep Learning for Computational Cognitive Modeling

Evidence accumulation models (EAMs) provide a powerful framework for inferring latent cognitive processes from choice and reaction time data. While EAMs are traditionally limited to binary choices, recent developments have extended them to rotationally symmetric response spaces via the circular diffusion model (Smith, 2016) and the spatially continuous diffusion model (Ratcliff, 2018). Yet, such extensions are limited in scope, as many psychological constructs are measured on bounded, non-rotational scales. In this paper, we bridge this gap by presenting and comparing two adaptations designed for bounded continuous data: the Half-Circular Diffusion Model (HCDM) and the Beta Drift Diffusion Model (BDDM). Because both models have intractable likelihoods, we fit them using Amortized Bayesian Inference (ABI) and compare them using Amortized Bayesian Model Comparison (ABMC). We demonstrate the complete workflow on an empirical affect dataset (N = 215), including parameter recovery, simulation-based calibration, posterior predictive checks, and model comparison. Both models accurately capture the joint distribution of responses and reaction times and yield interpretable parameters that can be reliably recovered. The model comparison further reveals a simple diagnostic for choosing between them: the dispersion of the rating distribution, with HCDM preferred for moderate spread and BDDM preferred for either highly concentrated or highly dispersed ratings. This work extends the EAM framework to a new application context of bounded continuous self-report data and offers researchers a user-friendly toolkit for modeling the cognitive dynamics of continuous responses. We release fully documented Python code with both GPU and CPU implementations, along with example datasets.

This is an in-person presentation on July 19, 2026 (10:40 ~ 11:00 EDT).

Trial-by-trial variability in cognitive processing may carry meaningful information beyond mean parameter estimates. In the Diffusion Decision Model (DDM), fluctuations in drift rate, boundary separation, and non-decision time can reflect dynamic adaptations of attention, strategy, and control. Yet estimating such dynamics—especially in individual-differences research with limited trial counts—poses serious challenges for conventional likelihood-based approaches. I utilize a simulation-based inference (SBI) framework for amortized Bayesian inference in both hierarchical and dynamic (superstatistical) DDMs. Neural posterior estimators are trained on large-scale simulations of the generative model, enabling efficient and stable recovery of mean parameters and structured trial-by-trial variability, including their temporal dependencies. Across multiple experimental paradigms, the recovered dynamic patterns differ substantially, and their relationships to cognitive abilities vary across tasks. These results highlight that within-person variability is shaped by task context rather than being a generic property of the model. Methodologically, the approach demonstrates how deep learning–based SBI makes it feasible to study individual differences in cognitive dynamics, opening new avenues for modeling complex trial-level processes.

This is an in-person presentation on July 19, 2026 (11:00 ~ 11:20 EDT).

In cognitive modelling, hierarchical Bayesian models are commonly used to capture individual differences while pooling information across participants, stabilizing estimates and improving generalizability and inference. Unfortunately, cognitive models are already computationally expensive, and extending them to large datasets with many participants dramatically exacerbates these costs. This is a challenge for MCMC-based approaches, where all individual data must be re-evaluated at every iteration, rendering large-scale hierarchical inference impractical. One approach to improving scalability is simulation-based inference (SBI), within which neural posterior estimation (NPE) is a commonly used approach for efficient Bayesian inference. NPE enables amortized inference: once trained, the estimator can be reused across individuals at negligible cost. However, NPE does not naturally extend to hierarchical models, as the dimensionality of the parameter space grows with each additional participant and simulation costs eventually outweigh amortization benefits. In this work, we build on recent methodological advances by Arruda et al. (2026) to enable scalable hierarchical Bayesian modelling. Rather than amortizing inference over the full hierarchical parameter space, we adopt a compositional approach that learns reusable, amortized update rules linking participant-level data to population-level parameters, which can then be composed across arbitrarily many participants. We demonstrate the effectiveness of this approach by fitting the diffusion decision model to a very large-scale dataset, showing that hierarchical cognitive modelling can be made computationally tractable without sacrificing its inferential benefits.

This is an in-person presentation on July 19, 2026 (11:20 ~ 11:40 EDT).

Many cognitive process models are straightforward to simulate but difficult to fit because their likelihoods are unavailable, analytically cumbersome, or too slow for repeated Bayesian inference. We evaluated whether approximate likelihoods made inference for such models more practical. We compared kernel density estimation (KDE), a direct non-parametric density approximation, with neural likelihood estimation (NLE), which we implemented using normalizing flows. We simulated from three models of increasing complexity: the log-normal race model, the racing diffusion model, and the diffusion decision model with across-trial variability. We sampled model parameters from broad, literature-informed ranges using Sobol sequences and split simulated datasets by parameter set into training, validation, and test sets. We evaluated performance at two levels. First, we compared approximate likelihoods with analytical likelihoods using error, divergence, and correlation metrics. Second, we embedded each approximation in MCMC to assess posterior bias, coverage, posterior predictive performance, and simulation-based calibration. We also compared simulation budget, training or fitting time, and inference speed. We found NLE to be more robust than KDE as dimensionality and model complexity increased, while KDE remained useful for simpler models or larger simulation budgets. We discuss practical guidance for selecting simulation-based likelihood tools in cognitive modelling, including cases where exact likelihoods existed but were too slow for routine Bayesian workflows.

This is an in-person presentation on July 19, 2026 (11:40 ~ 12:00 EDT).