DynVision: A Toolbox for Biologically Plausible Recurrent Convolutional Networks

Convolutional Neural Networks (CNNs) have demonstrated remarkable success in image recognition and exhibit conceptual similarities to the primate ventral visual pathway. Adding recurrence opens the door to exploring temporal dynamics and investigating mechanisms underlying recognition robustness, attentional modulation, and rhythmic perception phenomena. However, modeling spatiotemporal dynamics of biological vision using CNN-based architectures remains challenging. Incorporating functionally beneficial recurrence, capturing biologically plausible temporal phenomena such as adaptation and subadditive temporal summation, and maintaining topographic organization aligned with cortical structure require significant computational considerations. Although recent advances have incorporated neurobiological constraints, the field lacks accessible tools for efficiently integrating, testing, and comparing these approaches. Here, we introduce DynVision, a modular toolbox for constructing and evaluating recurrent convolutional neural networks (RCNNs) with biologically inspired dynamics. Our approach facilitates the incorporation of key visual cortex properties, including realistic recurrent architectures, activity evolution governed by dynamical systems equations, and structured connectivity reflecting cortical arrangements, while maintaining computational efficiency. We demonstrate the frameworks utility through systematic analysis of emergent neural dynamics, highlighting how different biologically motivated modifications shape scientifically-relevant response patterns. Code can be found at: https://github.com/Lindsay-Lab/DynVision/