The standard view of how we make sense of the world around us focuses on reconstructing our environment from the information received by our sensory organs. In this view, low-level brain areas (e.g., primary sensory cortex) represent basic features of objects, which are elaborated on in successive processing stages, until representations become increasingly complex in high-level areas (e.g., frontal cortex). An alternative view is predictive coding (PC), in which we model our environment to generate sensory predictions. In PC, high-level brain areas generate predictions of sensory activity and transmit them to low-level areas. A prediction that does not match the sensory information gives rise to a prediction error. This error signal is sent from low- to high-level brain areas to update the model of our environment, thereby improving future predictions to minimize errors. Modeling studies show PC is a fast and efficient way to process sensory information, and PC provides innovative hypotheses for understanding sleep and anesthesia, particularly when disconnected consciousness occurs (consciousness without awareness of the environment), like dreaming. PC also holds great promise for conceptualizing and treating brain disorders, including schizophrenia and depression. But key central features of PC have not been empirically tested and little is known about the underlying neural mechanisms. The goal of the proposed project is to characterize the neural dynamics, circuits and receptors enabling PC. There are two principle hypotheses. First, predictions depend on N-methyl-D-aspartate receptors (NMDAR) because NMDAR influence the activity of high-level brain areas where predictions are generated, and NMDAR are enriched on neurons in lower-level areas receiving predictions. Second, in disconnected consciousness, a breakdown of information transmission from low-level to high-level brain areas, as well as a breakdown of computations within each area, explains why models of our environment are not updated by external sensory information. These breakdowns prevent the comparison of predictions and sensory information, as well as the transmission of prediction errors to high-level brain areas. To test these hypotheses, we use a cross-species experimental design connecting cellular, circuit and systems levels to behavior. We will perform electroencephalography, machine learning and computational modeling to define the neural basis of PC in humans performing prediction tasks. Then we will manipulate PC using different anesthetic agents with diverse mechanisms, establishing causal relationships between receptors, large-scale brain networks and PC. In parallel, we will simultaneously record activity from sensory and high-level brain areas of non-human primates (NHPs) using the same PC tasks and pharmacological interventions to measure cellular and circuit level contributions to PC. Investigating PC will illuminate the fundamental mechanisms of perception, providing critical insights to guide therapeutic development for multiple health conditions.
This research advances our understanding of how the brain generates predictions which shape how we perceive the world, otherwise known as predictive coding. Predictive coding is relevant to public health because it holds great promise in conceptualizing and treating brain disorders, including depression, schizophrenia, delirium and dementia, as well as understanding sleep and anesthesia. Progress in understanding the basic mechanisms of predictive coding, as well as predictive coding deficits, is a necessary step in developing more effective treatment strategies for multiple health conditions.
