Predictive Processing, Hierarchical Predictive Coding, and Multi-Modal Evidence from Local/Global Oddball Paradigms

Abstract

Prediction is thought to be a fundamental function of the cortex. Before the adoption of predictive frameworks and invention of predictive paradigms in psychological and neuroscientific research, many studies in the field addressed concepts related to prediction. These include habituation, conditioning, and attention. The discovery of mismatch negativity marked the advent of many predictive paradigms, most prominently “oddball paradigms”. On the side of computational modeling, many have drawn inferences from neuronal mechanisms and the hierarchical and recurrent (i.e. involving feedforward/feedback streams) nature of the brain to successfully simulate neuronal behavior when it comes to prediction and prediction error. This also inspired theoretical works such as the Bayesian brain hypothesis and the free energy principle, which mathematically formalized the brain’s mechanism to minimize prediction error. An important paradigm that arose from the study of hierarchy and predictive processing was the local/global oddball paradigm. In this paradigm, repetition-driven predictions are orthogonalized with pattern-driven predictions, allowing for the delineation of brain areas and mechanisms of two different levels of predictive processing. Studies using fMRI, EEG, and LFP somewhat consistently found that local oddball effects occur earlier, are found in sensory and higher order cortical areas, and are related to gamma power increase. Global oddball effects, on the other hand, occur later, are found predominantly in frontal areas, and are associated with alpha/beta power decrease. However, studies examining single-unit spiking were inconsistent in showing significant global oddball encoding in the cortex. This may be due to the anesthetized state of subjects, lack of learning, and under-sampling of brain areas and/or neurons in a cortical column. Further studies using multi-area high-density recording methods in awake, behaving subjects are needed to elucidate cellular level predictive mechanisms.