Abstract
We regularly encounter the same objects, people, and places in predictable spatial and temporal configurations (e.g., a sequence of streets and landmarks on a daily commute). The human brain is highly attuned to this structure, as evidenced by the rapid extraction of simple regularities in studies of statistical learning. However, real-world regularities are more complex, often clouded by idiosyncrasies across repetitions (e.g., variable traffic patterns, music playing, and weather), requiring generalization to uncover higher-order structure. Little is known about how the brain extracts and represents structure at various levels of abstraction. To address this question, we recorded intracranial EEG from epilepsy patients while they viewed a stream of scenes containing different levels of regularities. In the exemplar-level condition, the same exact photographs were presented in repeating pairs (e.g., scene B always followed scene A). In the category-level condition, scene photographs were trial-unique but categories were paired (e.g., a mountain always followed a beach). We used a technique known as frequency tagging (which capitalizes on neural entrainment to periodic stimuli) to detect statistical learning in each condition, relative to a random control condition. Throughout the brain, we found robust frequency tagging not only to individual stimuli but also to learned pairs in both the exemplar-level and category-level conditions. The category-level result suggests that the brain abstracts over trial-level variance online to represent higher-order regularities. We next tested whether the same neural sites were involved in learning both exemplar- and category-level structure or whether these two levels of learning recruit distinct neural mechanisms. Although there was reliable overlap in the electrodes exhibiting both kinds of learning, there were also electrodes that represented only category-level or only exemplar-level regularities. Together, these findings provide initial insight into how the brain supports flexible and robust statistical learning across the visual hierarchy.