Abstract
We live in a highly structured world, repeatedly encountering the same objects, people, and places in a reliable fashion. Our mind is deftly attuned to such structure, quickly extracting spatial and temporal regularities via statistical learning. Yet in natural settings, even regular visual input is not identical across repeated experiences and we must learn to extract stable properties across noisy, idiosyncratic instances. Behavioral studies have provided evidence that statistical learning can abstract over such noise to represent high-level regularities. However, these studies relied on separate tests after exposure, raising the possibility that some abstraction may occur through inference across instances at test rather than through online integration during statistical learning. Here we leveraged the spatiotemporal resolution of human intracranial EEG to test whether category-level regularities can be learned online. Patients with epilepsy were exposed to a rapid continuous stream of scene images that were all trial-unique. In the Structured block, the categories of these scenes were paired (e.g., beach-mountain), whereas in the Random block, scenes from other categories were presented in a random order, preventing category-level regularities. Using frequency tagging, we found robust synchrony at the frequency of the image presentation, reflecting neural entrainment to visual stimuli. In the Structured but not Random block there was additionally synchrony at half of the image frequency, reflecting the (learned) boundaries of the category pairs. Critically, this category pair synchrony emerged even though the images differed at the exemplar level, providing online evidence for category-level statistical learning. Using multivariate pattern classification, we further found that paired categories came to be represented more similarly even when presented individually. Specifically, images from the first category in a pair resulted in neural evidence for both the first and second category. These data provide insight into the dynamics and representational changes underlying abstraction during statistical learning.