A key cognitive function is expectation. Expectation is thought to be generated through an agent’s experiences and learning. An established theoretical model, predictive coding, states that the brain is constantly building models (signifying changing expectations) of the environment. The brain does this by forming predictions (PD). These predictions interact with incoming sensory data. When the PD matches the sensed data, the expectation is correct. When they do not match, a prediction error (PE) signal is generated. This PE signal is then used to update the prediction, so that the brain’s internal model can more optimally predict future sensory data. The implications for the predictive coding model are far-reaching. If the model is correct, it would fundamentally shift our understanding of the neural code from one that represents the “state of the environment” (e.g., the classic Hubel and Wiesel receptive field model) to one in which the brain performs “active sensing” and builds internal models of the world, testing them against incoming sensory data. In addition, the predictive coding model has many implications for our understanding of disease states. For example, autism can be understood as a failure in correctly predicting social actions, and as a result, every social interaction is “surprising”. Various theories exist about how a predictive code could be implemented in the brain. They propose that distinct cortical layers, flow of communication (feedforward/feedback), and oscillatory dynamics are involved in signaling PEs and PDs. However, little neurophysiological data exist to support these models. Here, I propose an experiment to manipulate predictions by changing the probabilities associated with objects in a delayed- match-to-sample task (Aim 1). This will allow me to induce expectations of varying strengths. With my primary mentor, Earl Miller, I will be trained to perform make multi-area, multi-laminar recordings in monkeys. I will then use these data to study how expectations are built and what happens when they are violated. In Aim 2, with my secondary mentor, Nancy Kopell, I will use computational modeling to understand how the changing probability of inputs map on to a synchronously firing co-active group of cells (an assembly). We hypothesize that different assemblies represent different predictions. We also hypothesize that the strength of each assembly will represent the probability of a particular stimulus (thereby forming the neural basis of PD). Finally, due to the excitatory-inhibitory loops between cells in an assembly, we will investigate whether re-activations of the assembly occur rhythmically, paced by a beta (15-30 Hz) oscillation in deep cortical layers. Gamma oscillations (40-90 Hz) in superficial cortical layers could help switch off the current prediction (PD) by signaling prediction error (PE). In Aim 3, we will test whether interrupting beta oscillations (thought to signal PD) with closed-loop optogenetic inhibition is sufficient to disrupt the behavioral and neuronal signatures of prediction. This combination of experiments and biophysical modeling is poised to significantly contribute to our understanding of an important theoretical model of brain function, predictive coding.

Predictive coding is a computational strategy for neural circuits to optimally represent the environment, and these coding mechanisms are thought to be impaired in many brain disorders, such as autism. This project will combine large-scale electrophysiology, computational modeling, and optogenetics to test various architectures that may be used in the brain to implement a predictive neural code. A thorough understanding of these mechanisms may lead to better insight into the causes of many disorders such as autism.