Geometric models of the representation in working memory
Date:
Abstract
A central issue in understanding what we can attend to and remember is determining how information is represented in mind. In this talk, I work through my nascent thinking on what defines a representation, through the lens of representational geometry in behaviour and in the brain. First, in behaviour: Classic models of cognition use multidimensional scaling of similarity judgements to determine a psychological representation (e.g. the Generalized Context Model; Nosofsky, 1986). A comparable model (the Target Confusability Competition Model; Schurgin, Wixted and Brady, 2020) has seen an influential rise in explaining working memory performance. I will present recent work that uses Bayesian MCMC methods to infer the representation underlying various cognitive tasks; this work brings into question whether similarity can be assumed to be the basis for visual cognition. Second, in the brain: A popular approach in cognitive neuroscience has been representational similarity analysis (RSA) – classifying the pattern of neuroimaging activity under varying conditions using machine learning. I will present recent work that uses this approach to decode working memory load following associative learning (“chunking”), showing how learning changes the representational geometry. I will end with some reflections on the concept of representations, and how that might affect the goal of “representational alignment” – bringing these representational geometric models together.
Geometric models of the representation in working memory by William Xiang Quan Ngiam is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.