How the visual brain can learn to parse images using a multiscale, incremental grouping process.

Area of Science:

• Computational neuroscience
• Computer vision
• Perceptual psychology

Background:

• Object-based attention groups image fragments belonging to the same object.
• Curve-tracing tasks assess the ability to group elements of elongated curves.
• Previous research suggests curve-tracing speed is influenced by the distance between curves.

Purpose of the Study:

• To propose a recurrent neural network architecture for scale-invariant tracing of curves and objects.
• To investigate the mechanisms underlying multiscale perceptual grouping in the brain.
• To provide a computational model that explains how the visual system learns and executes curve tracing.

Main Methods:

• Developed a recurrent neural network with feedforward and recurrent pathways.
• Incorporated a disinhibitory loop involving VIP and SOM interneurons for activity propagation.
• Trained the network using a biologically plausible reinforcement learning scheme.

Main Results:

• The network demonstrated generalization from short to long curves and 2D objects.
• The model dynamically selected tracing scale based on curve distance and object width.
• Network behavior mirrored human psychophysics and monkey visual cortex findings.

Conclusions:

• The proposed architecture offers a mechanistic account of multiscale perceptual grouping.
• The findings provide insights into the neural basis of attention and visual processing.
• The model successfully integrates bottom-up and top-down processing for visual segmentation.