Modeling visual working memory using recurrent on-center off-surround neural network with distance dependent inhibition.

This paper presents a computational model of visual working memory (VWM) that simulates the processing of spatially distributed objects and their features. The model emphasizes the prioritization of object-related information before feature-related processing, effectively reversing the conventional feedforward order in visual perception. We conduct a detailed stability analysis to demonstrate non-divergence through energy function evaluations, highlighting the robustness of the network under varying input conditions. Additionally, we investigate the model's performance in change detection tasks through error rate analysis, focusing on how receptive field sizes and input configurations (crowded or spaced out) affect accuracy. This paper further extends the distance dependent RNN idea to a hierarchical computational model of visual working memory designed to parallel the structure and function of the visual processing hierarchy (VH). The model introduces a top-down feedback mechanism that dynamically refines object and feature representations, prioritizing object-level clarity initially and progressively localizing feature details through recurrent processing. A novel cross-talk function modulates feature specificity based on feedback depth, capturing the uncertainty in representations before full localization reaches lower visual areas.