Neurocognitive modeling

Traditionally, processing stages were investigated using behavioral measurements. To better capture the ongoing process, researchers have recently turned to neuroimaging methods instead. In that frame, a novel machine learning algorithm, hidden semi-Markov model multivariate pattern analysis was introduced (HsMM-MVPA; Anderson, Zhang, Borst, & Walsh, 2016). The goal of the current project was to validate HsMM-MVPA as a method for discovering stages directly from EEG data. To that end, two simple visual discrimination tasks were designed in which perceptual processing and decision difficulty were manipulated. For comparison with standard RT-based methods, the HsMM-MVPA analysis was complemented with evidence accumulation models (EAMs). The results of the analysis revealed that five-state HsMMs accounted for the data in all tasks. The brain activation of one of those stages was dependent on perceptual processing, while the brain activation and the duration of two other stages were dependent on decision difficulty. Consistent with the HsMM-MVPA results, EAMs showed that non-decision time varied with perceptual difficulty and drift rate value with decision difficulty, respectively. Additionally, non-decision and decision time of the EAMs correlated highly with the first two and the last three stages of the HsMM-MVPA analysis, respectively, indicating that the HsMM-MVPA analysis gives a more detailed description of stages discovered with this more classical method. Our conclusion is, therefore, that cognitive stages can be directly inferred from EEG data with the HsMM-MVPA analysis.

Cognitive control processes play an important role in many substantial psychological theories, but are hard to reliably and validly measure on the subject-level (Hedge et al., 2018; Rouder et al., 2019). Therefore, associations between individual differences in cognitive control and other variables are often inconsistent. Here we present a model-based cognitive neuroscience approach of cognitive control in which we integrated a mathematical model– the dual-stage two-phase model (Hübner et al., 2010) – with electrophysiological correlates of selective attention. We analyzed data from 149 participants who completed the Eriksen Flanker task while their EEG was recorded. We used structural equation modeling to a) improve the reliability and precision subject-level estimates by modeling them on a latent level and b) directly test competing theoretical higher-order linking structures between model estimates and latencies of the lateralized readiness potential. We will demonstrate that model parameters and neural correlates showed convergent validity and could be meaningfully related to each other. Together, these neurocognitive process parameters jointly predicted 37 % of the variance in individual differences in higher-order cognitive abilities. We propose that model-based cognitive neuroscience approaches can be used to overcome the measurement crisis of individual differences in cognitive control.

The neural mechanisms underlying attention-based perceptual decisions are of vital importance to a comprehensive understanding of behavior and cognition. Recent work has suggested that attention may play a key role in perceptual decision making. However, the exact cognitive components involved as well as the biomarkers of attention to predict behavioral performance in perceptual decisions have not yet been determined. To accomplish this, based on the Bayesian hierarchical diffusion model we have explored the underlying latent process of spatial attention in perceptual decision processes simultaneously at the group and individual level. The model’s parameters discovery showed that non-decision time (encoding plus motor execution) received the smallest deviance information criterion (DIC) and largest R-square relating to prioritized and non-prioritized top-down spatial attention. Moreover, based on the event-related potential (ERP) analysis and multiple linear regression model, N2 sub-component contralateral amplitude at central electrodes and alpha power band at parietal-occipital can predict very well response time (RT) relating to to-down spatial prioritization. But, the non-decision time parameter was predicted by only the contralateral N2 sub-component and not contralateral alpha power. Conversely, ipsilateral N2 sub-component and alpha power could not interpret the modulation of spatial prioritization in the decision process. In order to verify the convergence of the Markov chain Monte Carlo (MCMC) sampling, the R-hat Gelman-Rubin statistic was under 1.0001 which appears that the best scenario of the diffusion model was superior convergence and the same stationary distribution.

In recent years, the cognitive neuroscience literature has come under criticism for containing many low-powered studies, limiting the ability to make reliable statistical inferences. Typically, the suggestion for increasing power is to collect more data with neural signals. However, many studies in cognitive neuroscience use parameters estimated from behavioral data in order to make inferences about neural signals (such as fMRI BOLD signal). In this paper, we explore how cognitive neuroscientists can learn more about their neuroimaging signal by collecting data on behavior alone. We demonstrate through simulation that knowing more about the marginal distribution of behavioral parameters can improve inferences about the mapping between cognitive processes and neural data. In realistic settings of the correlation between cognitive and neural parameters, additional behavioral data can lead to the same improvement in the precision of inferences more cheaply and easily than collecting additional data from subjects in a neuroimaging study. This means that when conducting an neuroimaging study, researchers now have two knobs to turn in a design analysis: the number of subjects collected in the scanner and the number of behavioral subjects collected outside the scanner (in the lab or online).

Fitting drift-diffusion models (DDMs) to multiple participants’ choices and response times during perceptual decision making tasks result in parameter estimates that have cognitive interpretations such as individual differences in speed-accuracy tradeoffs and the average rates of evidence accumulation. The cognitive interpretations of DDM parameters can then be verified with experimental conditions and manipulations. Fitting neural drift-diffusion models (NDDMs) to participants’ scalp-recorded EEG as well as choices and response times can further reveal additional individual differences in cognition, such as individual differences in visual attention, visual encoding time (VET), and evidence accumulation processes. We discuss our recent efforts to develop NDDMs that are useful in understanding differences across individuals. In particular we are interested in models that actually recover parameters from simulated behavior and EEG data. Often newly developed NDDMs converge to a solution when using hierarchical Bayesian methods. However, whether the posterior distributions of parameters are informative about individual differences is not clear unless parameter recovery and parameter generalization to similar models are confirmed. In particular we discuss modeling efforts to understand individual differences in cognition that cannot be learned with models of either EEG or behavior alone.

Understanding the decision-making process is crucial to any theory of cognition. A popular framework for the mathematical modeling of decision-making is the sequential sampling framework. Support for this framework comes from converging evidence from animal studies showing the implementation of processes similar to evidence accumulation in several brain regions. While there is continued debate about which brain regions play critical roles in the perceptual decision-making process, several recent studies suggest the superior colliculus (SC) is involved. In one such study, rhesus monkeys completed a simple perceptual decision-making task with and without inactivation of neurons in the intermediate layers of the SC via muscimol injection. The monkeys made fewer responses to targets presented in the inactivated receptive field and the correct responses made towards the inactivated field were slower than in the pre-inactivation condition. Previous work found that a Diffusion Decision Model (DDM) allowing the drift rate parameters to vary across the injection conditions was the preferred model for these data, implying that the inactivation of the SC affected the rate of evidence accumulation. Since muscimol is a GABA agonist and there are GABAergic neurons in the SC, it is possible that the muscimol inactivation affected the competitive dynamics instead of simply the drift rate. Subsequently, we build upon the prior work by fitting (in addition to the DDM) two models that instantiate competition, or the lack thereof, differently than the DDM: the race model and Leaky Competing Accumulator (LCA) model. When fitting to the data, we allowed either the drift rate, decision threshold, neural leak, or lateral inhibition to vary across the pre-inactivation and post-inactivation conditions. Regardless of which parameter was manipulated across conditions, the LCA models provided a better fit to more sessions than the DDM or race models. The two winning models were the LCA model where the drift rates decreased in the post-inactivation condition relative to the pre-inactivation condition, and the LCA model where the neural leak increased in the post-inactivation condition relative to the pre-inactivation condition. Our modeling results provide further evidence that the SC is involved in decision-making, and that interactive competition plays a key role in the dynamics of the accumulation process.