Neurocomputational Models of Task Representation

Human cognition is purposive. We can set our minds toward attaining a particular desired goal, ignoring distractions and overriding counterproductive habits. Yet, it is also multipurpose. We are not permanently bound to pursue a singular goal, or even a small number. As drives and circumstances change, we can flexibly set our minds to attaining more suitable ones. How does the brain achieve this “setting” and “resetting” — this capacity for cognitive control?Increasingly, cognitive neuroscientists are focusing on studying representations that underlie cognitive control, formalizing their ideas within neurocomputational models (Botvinick and Cohen, 2014). Two distinct models of task representations have been highly influential: Guided Activation (Miller and Cohen, 2001), and Adaptive Coding (Duncan, 2001) through Random Mixed Selectivity (Rigotti et al., 2013). While these models have several commonalities --- both posit that the prefrontal cortex (PFC) drives controlled behavior through task representations --- they differ greatly in the representations they posit, and more generally reflect discrepant philosophies of understanding brain function.Here, we provide an overview of cognitive research into cognitive control, focusing on the role of task representations. A detailed examination of the Guided Activation and Adaptive Coding or Random Mixed Selectivity models is provided, in which models are evaluated in relation to the extant experimental literature and compared with respect to broader issues regarding task representations.