Search Papers | Poster Sessions | All Posters
Poster B86 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Spatial computing: utilizing cortical space to dynamically control the cognitive status of neural representations
Abhirup Brandyopadhyay1 (), Pawel Herman2, Mikael Lundqvist1; 1Karolinska Institute, Sweden, 2KTH Royal Institute of Technology, Sweden
Working memory (WM) is the ability to transiently store and selectively control a limited set of information in support of higher-order cognition. However, the neural basis of this flexible control is disputed. We recently proposed that cortical network space might be utilized to dynamically update the cognitive status of memory representations and provided experimental evidence for this principle, referred to as spatial computing. Here we implement spatial computing in neural mass models where space is explicitly represented to test its computational feasibility, structural requirements, and generalization capabilities. We demonstrate that distant-dependent like-to-like connectivity and local winner-takes-all-dynamics, both observed in the cortex, are sufficient requirements. In our implementation, spatio-temporal patterns of externally-imposed inhibition dictate where and when information is stored. Thus distinct memory items are encoded in distinct spatial locations, enabling the network to implement selective task-dependent control of WM-representations. The cortical locations of an item can be dynamically updated as their cognitive status change. The imposed, task-dependent inhibition yields a low-dimensional activity pattern independent of item-specific WM information, thus cognitive control generalizes to new patterns. Further, spatial computing can be implemented in WM-networks relying on either persistent activity or synaptic mechanisms. Synaptic mechanisms facilitate intermittent activation of stored items. By imposing a travelling wave of top-down disinhibition distinct items, stored in distinct parts of the network, are activated at different phases of the travelling wave. Together, we demonstrate that low-dimensional dynamics based on utilizing the spatial dimension of cortical space as an information encoding dimension facilitates flexible WM-control and generalization.
Keywords: Cognitive control Spatial computing Generalization Neural mass model