Title Information
Type of Resource (primo): dissertations
Name: Personal
Name: Personal
Name: Personal
Name: Personal
Name: Personal
Name: Personal
Name: Corporate
Origin Information
Physical Description
Note: thesis: Thesis (Ph. D.)--Brown University, 2025
Genre (aat): theses
Abstract: Working memory is necessary for many everyday tasks and yet there are many open questions about working memory capacity, individual differences in working memory, and how learning interacts with working memory. The existing definition of working memory is a temporary, capacity-limited, storage of information that is intended to be used soon. Here, I discuss other key components of working memory, namely 1) input gating: selecting the information that goes into memory and how this information will be organized with respect to all other information in working memory, 2) encoding: how information encoding might change based on load, and 3) output gating: accessing the information stored in working memory. I create a biologically plausible neural network of the prefrontal cortex and basal ganglia, augmented with the ability to compress information during encoding (chunking). Model performance and internal gating strategies show that adaptive chunking can be learned via reinforcement learning, suggesting mechanisms for how this chunking occurs in the brain. Model comparisons show that managing working memory resources is critical to performance on working memory task – it is not enough to have the resources, but just as important to learn how to utilize them, a challenging feat in part due to the credit assignment learning problem. This framework makes novel predictions about working memory capacity differences in individuals and cognitive deficits in patient populations. Lastly, I test key computational principles taken from biological systems in artificial systems to better understand artificial systems as well as to improve them. Specifically, due to the nature of the input, Transformers have no working memory storage demands, and yet they must solve what to attend to and when. Experiments with transformers show that they also face the credit assignment problem similar to biological systems. I further show that mechanistic pretraining inspired by input and output gating from biological models can improve generalization and accuracy in transformers. Through the analysis of biological and artificial systems, this work provides a better understanding of the gating operations that are key components of working memory, the importance of learning how to utilize working memory resources, and individual differences in working memory.
Subject (fast) (authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/01732553")
Subject (fast) (authorityURI="http://id.worldcat.org/fast", valueURI="http://id.worldcat.org/fast/00872004")
Subject
Subject
Subject
Subject
Language
Record Information