Modeling working memory: Varying noise based on neural differences.
Open Access
- Author:
- Delmar, Sarah
- Graduate Program:
- Neuroscience
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- August 21, 2024
- Committee Members:
- Sonia Cavigelli, Program Head/Chair
Frank Ritter, Chair & Dissertation Advisor
Bradley Wyble, Outside Field Member
Nancy Dennis, Major Field Member
Farnaz Tehranchi, Outside Unit Member - Keywords:
- Cognitive modeling
Neuroimaging
Working memory - Abstract:
- While the neural mechanisms underlying learning and memory have been studied for decades, the details of how information is unconsciously encoded so efficiently during implicit learning have remained elusive. Implicit learning is prevalent everywhere; individuals learn how specific items or events fit into a sequence in complex tasks involved in life every day. Additionally, humans are very good at implicit learning without any instruction or training (Batterink et al., 2019; Nissen & Bullemer, 1987; Robertson, 2007). Infants learn the phonemes in their native language very early in life, without any supervised learning, as one well established example. Neuroimaging has allowed researchers to better understand how elements of implicit learning are encoded and represented, but the specific mechanisms that lead to this learning have not been widely established. Competing theories exist, but their validation has been difficult for implicit learning tasks because behavioral measures of unseen processes are not always consistent. Researchers have looked towards computational models to better understand the underlying mechanisms of these tasks. Cognitive models have been used to accurately generate predictions of human behavior for a variety of cognitive tasks. The goal of using a cognitive model is to better understand the cognition that occurs during a task as well as to produce human-like data in greater quantities than are normally collected from human participants. The equations that are developed that underly these models can be implemented in simpler models as a more basic, straightforward way to generate behavioral data for a task. Such a model is rooted in the theory that is well established and utilized by a community of researchers, but the model does not have the robustness of an actual cognitive model. Using fMRI data to establish the timing of such a model should result in higher accuracy when generating predictions of human behavior on a task, as well as provide insights of the neural mechanisms underlying similar tasks. This thesis creates a simpler, easy-to-use, R-based model that uses equations from the activation learning system within the ACT-R architecture to model the working memory processing that occurs during a sequence learning task. Multivariate analysis of neuroimaging data is used to shape the model to produce more human-like data. The use of neuroimaging results in combination with this model provides a methodology for model validation that is not currently a widely used method. Specifically, the results from the distinctiveness analysis are used to influence tuning of the error parameter. The modeling of cognitive functions could continue to improve with the use of this method in the future, because it incorporates human neural data from the task being modeled. Methods like this, that combine computational modeling and neural data, are a natural way forward in the field of cognitive modeling that should lead to more realistic models.