abstract Clay Holroyd

Clay Holroyd (University of Victoria, Canada)

Toward a unified model of anterior cingulate cortex function

The anterior cingulate cortex (ACC) has been implicated in a wide range of seemingly irreconcilable cognitive functions. Recently we proposed that many of these functions can be captured by assuming that ACC motivates extended behaviors to achieve larger task goals (Holroyd & Yeung, 2012). Couched within the formal theoretical framework of hierarchical reinforcement learning (HRL), this theory of ACC function holds that ACC exploits computational efficiencies afforded by collections of actions that are represented together based on their conjoint goals, called "options" in the language of HRL, which allow for behavior to be manipulated at higher levels of temporal abstraction (Botvinick, Niv & Barto, 2009). Here I will present 2 computational models that make explicit the underlying assumptions of theory while demonstrating its internal consistency. First, I will present a computational model based on principles of reinforcement learning that simulate the effects of ACC lesions on rat behavior in a variety of decision-making tasks involving effortful control (Holroyd & McClure, 2015). Second, I will illustrate how recurrent neural networks trained to predict extended sequences of behavior exhibit properties that are broadly consistent with distributed patterns of neural activity in rat ACC, and with event-related brain potential and functional hemodynamic measures of ACC function in humans (Shahnazian & Holroyd, in preparation). These simulations suggest avenues toward the development of a unified model of ACC function.