Luca Kämmer^1,2,3, Carmen Amme⁴, Philip Sulewski⁴, Martin N Hebart^1,2, Tim C Kietzmann⁴; ¹Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, ²Department of Medicine, Justus Liebig University, Giessen, ³Department of Psychology, Humboldt-University Berlin, ⁴Institute of Cognitive Science, University of Osnabrück

How we make sense of our visual world has largely been studied using paradigms in which people are presented with images while maintaining stable fixation. Yet natural vision is dominated by frequent saccades that bring relevant information into foveal view. Psychophysical work shows that perception changes when we actively move our eyes, but it remains unclear how eye movements reshape visual information processing in the brain. To uncover how incoming visual information is processed before and after fixation onset, we used the AVS dataset, a large-scale MEG and eye-tracking dataset of 5 participants free-viewing 4,080 natural images. To differentiate between low- and high-level visual features, we used the SPoSE dimensions, which are behaviourally derived dimensions that have been shown to account for visually evoked activation in fMRI, EEG, and MEG data. To isolate the effect that the dimensional loadings of each fixation patch have on neural responses over time, we combined them with the Unfold method, which deconvolves neural responses across fixations. Using this approach, we found that neural responses prior to fixation onset were modulated by high-level, but not low-level dimensions. Using beamformer source reconstruction, we also found that this predictive activation is present primarily in hippocampal and medial temporal regions, but not in early visual cortex. These findings suggest that during natural vision, humans actively predict incoming high-level features of saccade targets to facilitate subsequent processing.