Our research focuses on the fundamental questions about how task-set –– the collection of cognitive control demands required to perform a task –– is learned, stored, retrieved and generalized to new tasks, contexts and experiences in both young and aging populations. The methods used include lab- and web-based behavioral testing, functional neuroimaging, computational modeling and virtual reality technology. Below are several recent research directions:
How expectations of cognitive control demand emerge from statistical learning
To simulate the learning of cognitive control demand through statistical learning, we combine reinforcement learning (RL) and Bayesian modeling. In so doing, we have begun to establish a computational model of flexible cognitive control that has the ability to flexibly adjust its reliance on remote and recent history of experienced cognitive control demand based on the volatility (i.e., rate of change) in the environment (Jiang, Heller, & Egner, 2014, Neurosci Biobehav Rev). From a theoretical stand point, this model unifies long- and short-term conflict adaptation processes, which are two classic phenomena of flexible cognitive control and whose coexistence previous models failed to account for. fMRI combined with model-guided multi-voxel pattern analysis (MVPA) revealed that the key components in the model –– the learning rate in RL and the prediction of control demand –– are mapped onto an insula-dorsal striatum network (Jiang, Beck, Heller, & Egner, 2015, Nat Commun). Future research will improve this computational model and apply the model to investigate aging effects on statistical learning of cognitive control demand.
The roles of the hippocampus in the learning of cognitive control demand
We examined the mechanisms underlying one-shot episodic learning of cognitive control demand (Jiang, Brashier, & Egner, 2015, J Neurosci). The results showed (a) a collaborative effort of the dorsal striatum and the hippocampus in binding multiple features of a previous trial to form an integrated episode, and (b) that repetition of any feature in the next trial triggered retrieval of all integrated features to modulate cognitive control and task performance. Ongoing studies are examining the contributions of the hippocampus to (a) the generalization of cognitive control demand through associative memories and (b) the encoding and retrieval of cognitive control demand through its association with spatial contexts.
How expectations of cognitive control demand regulate cognitive control
Through a combination of behavior, computational modeling, and fMRI, our research indicates that proactive and reactive control map well onto the two key components in RL –– prediction and prediction error (i.e., the discrepancy between predicted and actual cognitive control demand), which respectively depend on the dorsolateral (dl) and dorsomedial (dm) prefrontal cortex (PFC). That is, prior to the detection of the cognitive control demand required by the present situation, proactive (anticipatory) cognitive control is driven by predicted demands for control. Supporting this view, our studies demonstrate that the encoding strength of the predicted control demand in dlPFC scales with the behavioral reliance on proactive cognitive control (Jiang, Beck, Heller, & Egner, 2015, Nat Commun). Furthermore, disrupting left dlPFC function using transcranial magnetic stimulation selectively removes proactive cognitive control from behavior (Muhle-Karbe, Jiang, & Egner, 2018, J Neurosci). The left dlPFC representation of predicted cognitive control demand also integrates predictions from multiple sources (Jiang, Wagner, & Egner, 2018, eLife), consistent with the finding that left dlPFC encodes domain-general cognitive control demand (Jiang & Egner, 2014, Cereb Cortex). Finally, once the cognitive control demand required by the situation is detected, dmPFC (including rostral anterior cingulate cortex) becomes engaged to reactively resolve the prediction error about control (Jiang, Beck, Heller, & Egner, 2015, Nat Commun; Jiang, Wagner, & Egner, 2018, eLife). Future research will aim to characterize how different learning systems (e.g., statistical learning and episodic learning) jointly guide flexible cognitive control.