Coordinated Drift of Image Representations in Human EEG

Sensory representations drift over time in many animal systems, but evidence in humans remains limited. Moreover, how perception remains stable despite representational drift remains unclear. Here, we used THINGS-EEG2, a dataset with multiple sessions of repeated presentations of natural images, to test whether image-evoked EEG responses drift across recording sessions and whether this drift is geometrically constrained. We found that image-specific EEG response patterns drifted in a manner that was coordinated across channels and concentrated along dimensions that encoded image identity. Although this impaired cross-session decoding accuracy on average, its effects were heterogeneous with some image pairs becoming harder to distinguish and others becoming more separable. Critically, global distance statistics and the dimensionality of the image response space were preserved across sessions. Simulations show that downstream Hebbian networks can adapt faster to structured drift that preserves dimensionality than when drift is unstructured. These findings suggest that human sensory representations can change over time while remaining amenable to rapid compensation via unsupervised learning mechanisms.