Task switching differentially perturbs neural geometry in the human frontal and temporal lobes

The process of switching between tasks is cognitively taxing, and gives rise to a task switching cost (TSC), an increase in reaction time immediately after switching [1]. While robust in humans, the presence of TSC in other animals is debated, being absent from some species entirely, thus necessitating study in humans [2,3]. The neurophysiological basis of TSC is also debated, with two competing explanations involving reconfiguration of medial frontal cortical (MFC) activity for the current task, or lack of inhibition of the previous task. To disambiguate between these explanations, we conducted single-unit recordings in epilepsy patients who were instructed to alternate between cognitive tasks. Our objective was to identify population-level neural signatures that explain TSC and arbitrate between the reconfiguration and inhibition models of task switching. We recorded 1023 neurons in 10 patients (15 sessions) from medial frontal and temporal structures. We found that MFC neurons form a strong hierarchical representation of task context, unlike neurons in the prefrontal cortex and temporal lobe. Furthermore, we identify two orthogonal context-encoding subspaces within MFC that are present during pre-stimulus baselines. One subspace emerges immediately after switching but is transiently present, and is not predictive of TSC magnitude. The other subspace persistently encodes task context. However, while switching, even after the instructions, this subspace retains a representation of the previous task, and updates to the current task after the first trial. The degree of previous-task representation in this persistent subspace after a switch is predictive of switching costs. These findings provide a population level explanation for both reconfiguration and inhibition, with reconfiguration entailing transfer of context information out of the transient and into the persistent subspace, and a lack of appropriate inhibition reflected in the previous-task-encoding in the persistent subspace after switching, which explains TSC magnitude.