Ahmet Ali Yaman¹ (), Daniel Kaiser^2,3,4,5, Nihan Alp¹; ¹Sabanci University, Istanbul, ²Neural Computation Group, Justus Liebig University Giessen, ³Center for Applied Computer Science and Data Science (ZAD), Justus Liebig University Giessen, ⁴Center for Mind, Brain and Behavior (CMBB), Universities of Giessen, Marburg, and Darmstadt, ⁵Cluster of Excellence “The Adaptive Mind”, Universities of Giessen, Marburg, and Darmstadt

Understanding how humans perceive scenes is essential for uncovering how the visual system interprets our surroundings. Prior works suggest that coherent scenes are processed more efficiently than distorted ones because their typical structure and content guide top-down predictions. Here, we aim to establish an objective neural marker that quantifies the integration of scene parts into a coherent whole. We utilize an EEG frequency-tagging paradigm to measure Steady-State Visual Evoked Potentials (SSVEPs). Twenty-five participants viewed 10-second flickering scene images. Images were presented in a 2×2 grid, with the top-left/bottom-right and top-right/bottom-left segments, flickering at different fundamental frequencies (6 Hz and 7.5 Hz). Critically, we manipulated the composition of the scene by re-combining parts from different scenes across the top-left/bottom-right and top-right/bottom-left segments. Scene parts could be (in)congruent in their semantic content (stemming from the same or different categories) or spatial structure (showing at their typical or atypical positions in the grid). We then analyzed fundamental frequencies, reflecting the processing of individual parts of a scene, and intermodulation (IM; e.g., 1.5 Hz) frequencies, reflecting nonlinear components that emerge when the brain integrates individual parts into coherently organized scenes. Our analysis focused on Signal-to-Noise Ratios (SNRs) of fundamental and IM frequencies from posterior-occipital electrodes. SNR values at fundamental frequencies did not differ across conditions, indicating that part-based computation was similar across conditions. In contrast, IM frequencies revealed a significant difference between conditions. Intact scenes elicited significantly stronger SNRs than all distorted variations except those with only categorical content distortion. Notably, the strongest decrease in IM frequency SNRs occurred in the spatially distorted conditions. These results suggest that the IM specifically reflects the binding of scene parts, a process heavily constrained by spatial structure. More generally, our findings demonstrate that SSVEP-based IM frequencies provide a robust and objective marker of scene integration.

Acknowledgements: N.A. and A.A.Y. are supported by the Scientific and Technological Research Council of Turkey (TUBITAK 1001) under the Grant Number 122K922. D.K. is supported by an ERC Starting Grant (PEP, ERC-2022-STG 101076057) and by the DFG (EXC 3066/1 “The Adaptive Mind”, project number 533717223).