Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Multi-regional neural network models capture distributed working memory representations in human fMRI

Eun Tack Cho1, Jacob A. Miller2, Daniel B. Ehrlich3, Nicole P. Santamauro4, Zailyn Tamayo4, Yvette Afriyie-Agyemang4, Youngsun Cho4, John D. Murray1,4; 1Dartmouth College, 2University of Miami, 3University of California, Berkeley, 4Yale University

Presenter: Eun Tack Cho

How does working memory flexibly maintain and guide action plans in dynamic environments? Recent work has proposed that "contingency representations"—mappings between expected future cues and actions—help unify these working memory and planning computations (Ehrlich & Murray, 2022). Yet, how these contingencies emerge across distributed cortical circuits to guide behavioral outputs remains unknown. To address this question, we trained human participants to perform the Conditional Delayed Logic (CDL) task while being scanned in fMRI. This revealed dissociable contributions of cortical networks during the delay: the sensory cue was maintained persistently in higher-order visual cortices, rules and contingency states were actively maintained across the fronto-parietal network, and specific motor outputs were maintained in the somatomotor cortex. To further examine how sensory representations are transformed into contingencies, we trained a biologically constrained, multi-regional recurrent neural network (RNN) model to perform the CDL task. The neural geometry of the RNN developed a functional hierarchy across regions, reminiscent of human delay-period activity. Perturbing task-relevant subspaces of the RNNs demonstrated that hierarchical partitioning of cue and contingency information confers robust protection against distractors. Together, these results provide a mechanistic account of how multi-regional neural circuits construct contingency representations to support robust, context-dependent behavior.

Topic Area: Memory, Learning & Knowledge Structures