Year: 2025
Type: preprint
Abstract: Working memory can hold many types of information and is crucial for cognition. Commonly, models of working memory maintain information such as hues or words by forming memory attractors through structured connectivity. However, real-world information can be novel, making it infeasible to use predefined attractors. In addition, most models—with or without attractors—have focused on maintaining binary categories instead of continuous information in each neuron. In the present study, we investigate how the brain might maintain working memory representations of arbitrary novel patterns with graded values. We propose an unstructured, rate-based network model in which each neuron has multiple dendrites. Each dendrite shows bistable activity, which qualitatively captures the conductance-based dynamics in the corresponding spiking model and emulates fast Hebbian plasticity. This network can flexibly maintain novel graded patterns under various perturbations without fine-tuning of parameters. Through analytical characterization of network dynamics during the encoding and memory periods, we identify different conditions that yield either perfect memories or several types of memory errors. Our analysis reveals a functional separation for network neurons into two groups with distinct behaviors. Overall, this architecture provides robust and analytically tractable storage of novel graded patterns in working memory.
Authors: Jiacheng Xu, Daniel L Cox, Steven J Luck
Institutions: Dartmouth College, University of California, Davis
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
Awards: G1021461588, G1523888516
OpenAlex ID: W4412360777