A Biologically Plausible, Dynamics-based Model of Memory Using Neurosymbolic Methods

Abstract

This thesis presents a novel model of memory that addresses previous gaps in working memory modelling. Traditionally, models of memory treat memory as an atemporal process; they present a vector or trace of a memory that only changes with item presentations. We instead provide a dynamics-based model of short-term memory using vector projection, and we further characterize its capacity and forgetting mechanisms using temporal decay and neural saturation. We also provide an auto-associative memory that uses bundles of high-dimensional vectors to implicitly associate items to each other, enabling bundle reconstruction from single components. Lastly, we modify the parameters of the short-term memory to create a primacy-biased integrator of task-contextual embeddings.

We validate our model and demonstrate its ability to generalize across domains by implementing it in a spiking neural network and simulating semantic and spatial memory tasks with minimal parameter changes between simulations. For each simulation, we represent stimuli (e.g., list items or spatial landmarks) and task embeddings (e.g., temporal position in a list or spatial position within a room) as Spatial Semantic Pointers, and we compare our model's results to human experimental results. Across the serial recall, free recall, and positional recall experiments, our model consistently demonstrates human-like performance, demonstrating no statistically significant difference for 72.5% of compared data-points and negligible-to-small effect sizes for 91.2% of compared data points. Across all data points, we find mean and median absolute effect sizes in the ranges of (0.125,0.203) and (0.082,0.189) respectively, suggesting significant coherence between model and experimental task results.