Development of localized oriented receptive fields by learning a translation-invariant code for natural images

Area of Science:

• Neuroscience
• Computational Vision
• Machine Learning

Background:

• Mammalian primary visual cortex neurons exhibit spatially localized, oriented receptive fields.
• Previous theories proposed sparse coding or independent component analysis for efficient image encoding.

Purpose of the Study:

• To investigate if transformation-invariant coding can explain the emergence of oriented receptive fields.
• To demonstrate a neural network model capable of simultaneous object recognition and pose estimation.

Main Methods:

• Employing a first-order Taylor series expansion for transformation-invariant image coding.
• Training cooperating neural networks for object identity ('what') and transformation ('where') estimation.
• Utilizing joint maximization of a posteriori probability for visual data generation.

Main Results:

• Learned receptive fields approximate localized first-order differential operators.
• Networks successfully factored retinal stimuli into object-centered features and invariant transformations.
• Experimental validation of the proposed coding strategy.

Conclusions:

• Transformation-invariant coding provides a viable explanation for learned receptive field properties.
• The developed neural network architecture effectively performs object and pose recognition.
• This framework offers insights into efficient visual information processing in biological and artificial systems.