Examining Extra-Classical Receptive Field Effects in the Long-Range Modulatory Feedback Model

Abstract

Visual perception research highlights the brain’s ability to dynamically interpret visual input through classical (CRFs) and extra-classical receptive fields (ECRFs). While CRFs focus on localized visual features, ECRFs provide contextual modulation critical for tasks like contrast normalization, figure-ground segregation, and boundary detection. Despite their efficacy in object recognition, feedforward deep neural networks (DNNs) lack mechanisms for context-dependent modulation, limiting their biological plausibility. This study investigates whether a biologically inspired long-range modulatory (LRM) feedback model exhibits ECRF-like properties observed in biological vision. Using visual stimuli directly adapted from Cavanaugh et al. (2002), which systematically varied in size, orientation, and spatial frequency, we analyzed kernel activations across two processing passes: Pass 1 (feedforward) and Pass 2 (feedback). Receptive field properties were quantified using metrics such as the Grating Summation Field (GSF) and Suppression Index (SI) to assess center-surround interactions. Results revealed that Pass 2 induced significant changes in spatial integration, including an expansion of the excitatory summation field (higher GSF values) and increased suppression effects in the first convolutional layer, resembling V1 feedback mechanisms. However, the second layer exhibited high variability, with kernels responding inconsistently to feedback, ranging from suppression to amplification, suggesting heterogeneous effects of modulation. Additionally, Pass 2 responses displayed variable participation across kernels, with nonzero activations distributed unevenly compared to Pass 1, which exhibited more uniform activation.