samimol/multiscale_tracing

This repository contains the code for training and evaluating a biologically inspired recurrent neural network that learns scale‐invariant curve tracing and object grouping. The model replicates key findings from human psychophysics and monkey visual cortex: it dynamically selects spatial scale and propagates enhanced activity through feedback and horizontal connections, enabling generalization from short curves to long contours and 2D shapes. The network is trained using a biologically plausible reinforcement learning rule and captures reaction time patterns seen in perceptual grouping tasks.