Task Preparation is Reflected in Neural State Space Dynamics

Poster Session A - Saturday, April 13, 2024, 2:30 – 4:30 pm EDT, Sheraton Hall ABC

Harrison Ritz¹ (hritz@princeton.edu), Aditi Jha¹, Nathaniel Daw¹, Jonathan Pillow¹, Jonathan Cohen¹; ¹Princeton University

We can flexibly reconfigure our neural information processing to accommodate a wide array of different tasks. These task switches are increasingly being modeled as dynamical transition between task representations, however the empirical evidence informing these theories is controversial or incomplete. Here, we explore how the putative cognitive dynamics during task switching are reproduced in whole-brain neural dynamics. To study whole-brain neural dynamics, we re-analyzed a recent EEG experiment on task-switching by Hall-McMaster and colleagues (2019). Before each trial, human participants were cued to whether they would need to respond to the color versus the shape of a compound stimulus (50% switch rate). Borrowing emerging methods from systems neuroscience, we fit ‘latent state space’ models to whole-brain EEG activity during this pre-trial epoch. These models infer the linear recurrent neural dynamics and task encoding that gives rise to multivariate timeseries of electrode voltages. Fitting this model to trial-level EEG (N=30), we found that our approach was highly accurate at predicting EEG activity on single held-out trials. Moreover, our state space models extended traditional multivariate encoding analyses by revealing rich interactions between task encoding and system recurrence. Standard ‘controllability’ analyses revealed that the spread of task information throughout the brain was elevated in conditions that placed demands on cognitive control (e.g., switch trials). This experiment provides preliminary support for the dynamical systems approach to modeling task reconfiguration. Moreover, it provides a new empirical basis for developing richer process models of how our brains support cognitive flexibility.

Topic Area: EXECUTIVE PROCESSES: Goal maintenance & switching