LGN-CNN: A biologically inspired CNN architecture.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Computer Vision

Background:

• The human visual system's early processing stages involve complex receptive fields.
• Convolutional Neural Networks (CNNs) are powerful tools for image analysis.
• Mimicking biological visual processing can enhance CNN capabilities.

Purpose of the Study:

• Introduce a novel biologically inspired CNN architecture, LGN-CNN.
• Investigate the functional properties of the LGN-CNN's first layer, inspired by the Lateral Geniculate Nucleus (LGN).
• Evaluate the model's performance in terms of rotation and contrast invariance and Retinex effects.

Main Methods:

• Designed a CNN with a first convolutional layer featuring a single filter approximating Laplacian of Gaussian (LoG) functions.
• The architecture is inspired by the receptive field profiles (RFPs) of LGN cells.
• Performed computational experiments to demonstrate rotation invariance and analyze contrast invariance and Retinex effects.

Main Results:

• The LGN-CNN's first layer exhibits rotational symmetry, approximating LoG functions and LGN RFPs.
• The model demonstrates rotation invariance and contrast invariance capabilities.
• Comparison with LoG functions shows similar Retinex effects, validating the biological analogy.

Conclusions:

• The LGN-CNN effectively approximates LGN and V1 receptive field profiles.
• The model replicates the Retinex effects observed in LGN cells' long-range connections.
• This biologically inspired approach offers a promising direction for advanced computer vision models.