Neurocognitive modeling

Humans have a limited amount of cognitive resources to process various cognitive operations at a given moment. Based on the Source of Activation Confusion (SAC) model of episodic memory, resources are consumed during each processing and once depleted they need time to recover gradually. This has been supported by a series of behavioral findings in the past. However, the neural substrate of the resources is not known. In the present study, over an EEG dataset of a free recall task, we identified a neural index reflecting the amount of cognitive resources available for forming new memory traces. We showed that consistent with the model predictions, the index was able to capture the sequential effect of word frequency and the primacy serial position effect. In addition, greater available resources at encoding, as characterized by the neural index, are associated with better memory at recall. This provides an alternative explanation for the subsequent memory effect (SMEs, i.e. differential neural encoding patterns between subsequently recalled versus subsequently non-recalled items), which has been previously associated with attention, fatigue and properties of the stimuli.

Traditional cognitive models, such as sequential sampling models (SSM), are unable to address some research questions about cognition and related neural mechanisms since they rely solely on behavior and lack explanations of brain data. Electroencephalography (EEG) can help a popular model of SSMs, the drift-diffusion model, to unravel latent processes that have not been previously discovered. For instance, although perceptual decision making often involves both decision processes and non-decision processes, researchers often seek to anatomize the subprocesses of decision variables while being less aware of the impact of some cognitive manipulations on the non-decision processes. We propose neuro-cognitive models that can utilize behavioral data and neural signals simultaneously to constrain two distinct groups of parameters as well as predict the onset time of evidence accumulation by single-trial EEG in occipito-parietal areas. Specifically, we strove to investigate a research question, why does performance in decision making improve when participants are exposed to top-down spatial orientation cues before the appearance of stimuli? Using a public dataset, we found evidence that spatial top-down attention can manipulate both visual encoding time (VET) and other non-decision times, but not decision variables such as evidence accumulation during a face/car perceptual decision-making task. To make this inference we used "directed'' neurocognitive models that have previously been used to find single-trial relationships between EEG measures and cognitive parameters. However we have also tested new "integrated'' models that predict EEG measures. These integrated neurocognitive models predict two sources of variance in EEG measures across single-trials: variance related to trial-by-trial changes in non-decision times versus measurement error in the observed EEG measures. We discuss how we fit these two classes of models and how they can be used to answer similar questions in future work.

Recently, cognitive modelers have become increasingly interested in supplementing behavioral data with neural or physiological measures. In order to complement approaches that use a generative cognitive model of behavioral choice data, we develop a generative model of modulations in the variance of the electromyographical (EMG) recordings associated with pressing one or two response buttons. This model provides estimates of key quantities of interest such as onset, offset, and amplitude of EMG bursts for each response. The hierarchical structure (i.e., trials nested within participants) yields group-level estimates for these parameters for each participant. We use particle Metropolis within Gibbs sampling (Gunawan et al., 2020) to efficiently obtain posterior samples from the model. The model can be used to address questions of interest about the EMG signal itself (such as between-condition differences) but also holds the promise of linking EMG parameters to cognitive model parameters in a joint model.

The diffusion decision model (DDM) has been successfully accounting for reaction times distribution and accuracy for simple choice tasks by assuming a diffusion evidence accumulation process. However, the model has been criticized for its ad hoc distributional assumptions of cross-trial variability in its parameters, such as the drift rate. Using a word recognition task with electroencephalography recording, the current study aims to include both exogenous factors (e.g., word frequency and study-test lag) and EEG signals corresponding to retrieval success as endogenous factors to account for trial-to-trial variability in drift rate. The EEG signal was calculated based on two classic components observed in recognition memory tasks that differentiate types of memory judgement, namely the frontal negative component (FN400) and the late positive component (LPC). As such, these EEG components were suggested to index memory processes. Our results have shown when assuming individual trial drift rate as a linear combination of these factors, better model fits were observed as indicated by superior model selection scores. While model fit did not improve for randomly selected EEG signals unrelated to the memory process, the benefit from the EEG endogenous factors (FN400 and LPC) further suggests an involvement of these EEG components in the evidence accumulation process.