Convolutional neural networks develop major organizational principles of early visual cortex when enhanced with...

Area of Science:

• Computational Neuroscience
• Visual System Modeling
• Artificial Intelligence

Background:

• Primate visual cortex displays organizational principles like cortical magnification and radial bias.
• These principles are thought to stem from retinal ganglion cell distribution and retinal-to-cortical convergence.
• Understanding these principles is key to deciphering visual processing.

Purpose of the Study:

• To investigate the origins of primate visual cortex organizational principles using computational models.
• To test the hypothesis that non-uniform retinal sampling and quasi-uniform convergence drive these principles.
• To explore the role of spatial frequency in radial bias.

Main Methods:

• Development of convolutional neural networks (CNNs) incorporating a novel retinal sampling layer.
• The retinal sampling layer simulates non-uniform retinal ganglion cell density.
• Analysis of organizational principles, including receptive field properties and radial bias, in the trained CNNs.

Main Results:

• CNNs with retinal sampling layers successfully replicated key visual cortex organizational principles.
• Radial bias was found to be dependent on spatial frequency, appearing only at high frequencies.
• Low spatial frequencies exhibited an orthogonal orientation bias, suggesting a novel origin for radial bias.

Conclusions:

• The interplay between non-uniform retinal sampling and quasi-uniform convergence can explain visual cortex organization.
• Radial bias is not a universal feature but is modulated by spatial frequency.
• Retinal sampling and convergence jointly determine spatial frequency resolution across the visual field.