Testing predictions of the Binding Pool model of visual working memory

Restricted (Penn State Only)
Author:
Swan, Garrett Scot
Graduate Program:
Psychology
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
May 11, 2017
Committee Members:
  • Bradley Paul Wyble, Dissertation Advisor
  • Bradley Paul Wyble, Committee Chair
  • Richard Alan Carlson, Committee Member
  • Frank Gerard Hillary, Committee Member
  • John C Collins, Outside Member
Keywords:
  • Visual working memory
  • computational model
  • Working memory capacity
  • Neural network
Abstract:
Throughout out daily lives, we use visual working memory (VWM) in both simple and complex tasks to briefly maintain visual information. How VWM works has been a hotly debated topic in psychology for many years. This thesis focuses on testing predictions generated from a computational model of VWM called the Binding Pool model. In this model, a memory is a set of binding pool nodes activated from concurrent activity of the visual features that compose an object and a token that individuates memories. The first prediction focused on how similarity affects memory precision. The model predicts that highly similar and dissimilar stimuli are more precise than stimuli with intermediate similarity. In two experiments, the results supported the prediction of the model. The second prediction focused trade-offs in retrieval precision as a function of the number of relevant features. The model predicts this trade-off should be present for multiple objects. In two experiments, participants reported the task irrelevant feature with coarse precision and were significantly less precise when the task irrelevant feature became relevant to retrieval, which support the prediction from the model. However, the ability to code features coarsely needed to be added to the model. The third prediction focused on how repetitions are encoded into memory. The model predicts that repetitions will be less precise than a single object. In two experiments, participants were significantly more precise with repetitions than a single object and random multiple objects. This result failed to support the model. A fourth prediction about the underpinning of the EEG correlates of VWM was also generated and tested using a delayed estimation task. The initial findings were mixed, but may provide insight into how the P3 components may relate to the disambiguation of targets from non-targets. As a result of these experiments, there were significant changes implemented to how the model encodes information. Taken together, these new results provide additional benchmarks for models of memory and helped the shape the architecture of the Binding Pool model to become a more general model of memory.