Representation of local geometry in the visual system.

Area of Science:

• Computational Neuroscience
• Computer Vision
• Differential Geometry

Background:

• The visual system processes complex information, including geometric features.
• Understanding the computational mechanisms of early visual processing is crucial.

Purpose of the Study:

• To demonstrate how receptive field (RF) profiles can compute exact partial derivatives of retinal illuminance.
• To establish a framework for understanding visual cortex units and geometric feature computation.

Main Methods:

• Utilizing convolutions with specific receptive field (RF) profiles to compute partial derivatives.
• Employing third-order jet extensions to create position-dependent geometries.
• Applying differential geometry principles to analyze visual routines.

Main Results:

• Convolutions with RF profiles precisely yield partial derivatives of blurred retinal illuminance.
• Arbitrary concatenations of RF profiles generate higher-order derivatives and increased blurring.
• A framework is established for "point processors" to compute geometric features like edge curvature.

Conclusions:

• The described method provides a novel way to understand and classify units in the primary visual cortex.
• This approach offers a new perspective on visual routines for computing geometric features using differential geometry.
• The equivalence between local jets and partial derivatives facilitates geometric computation.