Simulation-Based Inference Symposium

Contaminant observations and outliers often cause problems when estimating the parameters of cognitive models, which are statistical models representing cognitive processes. In this study, we test and improve the robustness of parameter estimation using amortized Bayesian inference (ABI) with neural networks. To this end, we conduct systematic analyses on a toy example and analyze both synthetic and real data using a popular cognitive model, the Drift Diffusion Models (DDM). First, we study the sensitivity of ABI to contaminants with tools from robust statistics: the empirical influence function and the breakdown point. Next, we propose a data augmentation or noise injection approach that incorporates a contamination distribution into the data-generating process during training. We examine several candidate distributions and evaluate their performance and cost in terms of accuracy and efficiency loss relative to a standard estimator. Introducing contaminants from a Cauchy distribution during training considerably increases the robustness of the neural density estimator as measured by bounded influence functions and a much higher breakdown point. Overall, the proposed method is straightforward and practical to implement and has a broad applicability in fields where outlier detection or removal is challenging.

Theories of dynamic decision-making are typically built on evidence accumulation, which is modeled using racing accumulators or diffusion models that track a shifting balance of support over time. However, these two types of models are only two special cases of a more general evidence accumulation process where options correspond to directions in an accumulation space. Using this generalized evidence accumulation approach as a starting point, I identify several ways to discriminate between absolute-evidence and relative-evidence models using various techniques from machine learning. First, an experimenter can quantify the information that decision makers considered to identify whether there is a filtering of near-zero evidence samples, which is characteristic of a relative-evidence decision rule (e.g., DDM). Second, a modeler can use machine learning to classify a set of data according to its generative model. Finally, machine learning can also be used to directly estimate the geometric relationships between choice options. I illustrate these different approaches by applying them to data from an orientation-discrimination task, showing converging conclusions across all three methods in favor of accumulator-based representations of evidence during choice. These tools can clearly delineate absolute-evidence and relative-evidence models, and should be useful for comparing many other types of decision theories.

In recent decades, evidence accumulation models—most notably the diffusion decision model (DDM)—have increasingly been used to investigate cognitive differences associated with aging. Notably, slower response times in older adults are often attributed not to reduced evidence accumulation speed but to prolonged non-decisional processes and increased decision caution, the latter reflected by the DDM’s boundary separation parameter. Older adults exhibit consistently higher boundary separation values, indicating they accumulate more information before making a decision. The Ornstein-Uhlenbeck (OU) model extends standard DDMs by incorporating a parameter for self-excitation (or leakage) in the evidence accumulation process. Strong experimental validations of the DDM have shown that adjustments in decision boundaries explain speed-accuracy tradeoffs. However, recent findings using the OU model suggest that variations in self-excitation may offer a more accurate account of these effects. This work investigates whether age-related differences in boundary separation persist in the OU model. We address this question by presenting findings from analyses of 18 experimental tasks and a large online dataset (N > 5,000,000). We perform Bayesian model comparison to determine whether the DDM or the OU model better accounts for the observed data, and examine how age patterns in cognitive parameters differ between the two frameworks. Our results suggest that differences in self-excitation across ages may at least partly explain the age-related effects previously attributed to boundary separations.

This symposium brings together recent advances from the vibrant intersection of deep learning, Bayesian inference, and computational cognitive models. The synthesis of these fields enables researchers to develop, fit, compare, and validate high-fidelity models of cognition which otherwise remain out of reach for standard statistical methods (e.g., maximum likelihood estimation of Markov Chain Monte Carlo). However, many aspects of this emerging new generation of computational tools remains underutilized in cognitive modeling. Thus, the symposium aims to highlight both methodological progress in the development of generative neural architectures for hyper efficient (aka amortized) statistical inference, as well as novel applications of these methods to substantive questions regarding the representation of cognitive processes. The methods part will focus on new approaches to robust amortized inference and new methods for comparing cognitive models on massive data sets. The symposium then highlights a few applications that benefit from simulation based inference approaches toward scientific progress. Topics will include an examination of how dynamic decision-making models can be compared using both posterior parameter estimation (using a free parameter to control the way evidence is represented) and model comparison using probabilistic classifiers. Another talk will investigate age differences in a standard diffusion decision model vs. the Ornstein-Uhlenbeck model. It will also include talks that highlight modular aspects of the simulation-based inference toolkit, including how frontier reinforcement learning models can be combined with arbitrary sequential sampling models to test novel mechanistic hypotheses concerning intertemporal (across trial) learning dynamics.

Neural amortized Bayesian inference (ABI) offers a computationally efficient solution to probabilistic inverse problems, but its robustness remains a critical challenge for real-world deployment. When confronted with observations that deviate from the training distribution due to model misspecification or domain shifts, neural posterior estimators can exhibit severe biases. We investigate two new robust approaches to improve the reliability of neural amortized Bayesian inference (ABI) under domain shifts and model misspecifications. The first approach leverages unsupervised domain adaptation (UDA) to align the embedding spaces of simulated and observed data, mitigating biases from unmodeled artifacts. However, we find that this alignment can lead to failures under prior misspecifications, raising concerns about its applicability under domain shifts. The second approach introduces a new strictly proper training objective based on Bayesian self-consistency. We systematically evaluate both methods across diverse misspecification scenarios, including high-dimensional settings, with a particular focus on their utility for cognitive modeling. Our results demonstrate that while UDA can be beneficial in certain cases, Bayesian self-consistency offers notable improvements in robustness even when observed data deviates substantially from both simulated and real training distributions. If these findings generalize across domains, they could mark a major step toward making neural ABI reliable for widespread application.

The cognitive neuroscience community has increasingly adopted sequential sampling models (SSMs) as choice functions in reinforcement learning (RLSSM), providing new insights into generative processes that govern decision dynamics within and across trials. This hybrid framework enables the study of previously undescribed cognitive mechanisms linking decision-making and learning. While prior accounts of choice reaction time (RT) effects have explained decision speed variations through Hick’s Law (as a function of choice set size), value-based differences, or interactions between learning and set size (i.e., the number of stimulus-response pairs to be learned), the influence of other cognitive processes involved in instrumental learning remains underexplored. Here, we propose and test the hypothesis that working memory (WM) load modulates speed-accuracy trade-offs during instrumental learning. Using the Reinforcement Learning – Working Memory (RLWM) task and model, which incorporates multiple interacting generative learning processes, we formally evaluate this hypothesis. The RLWM model integrates a sequential sampling decision process to jointly account for choice and RT distributions. Hierarchical Bayesian model-fitting results support our hypothesis, revealing that participants proactively and adaptively adjust decision boundaries as a function of WM load—specifically, increasing decision boundaries under higher WM load. Enabled by recent advances in simulation-based inference, identifying such neurocognitive mechanisms provides a powerful framework for characterizing cognitive function in both healthy and clinical populations.