How Convolutional Neural Network Architecture Biases Learned Opponency and Color Tuning

Area of Science:

• Computational Neuroscience
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are key in AI, but understanding their learned representations is challenging.
• Architectural changes, like adding bottlenecks, can impact CNN function.
• Existing methods from neuroscience offer quantitative ways to analyze visual systems.

Purpose of the Study:

• To develop methods for quantitatively comparing trained CNNs.
• To classify convolutional neurons based on spatial and color opponency.
• To investigate how CNN architecture, specifically bottlenecks, affects learned representations.

Main Methods:

• Adapted electrophysiology and psychophysics methods to derive spatial and color tuning curves for convolutional neurons.
• Classified neurons in CNNs with varying depths and bottleneck widths.
• Developed a hue sensitivity curve method for CNNs.

Main Results:

• CNNs with bottlenecks show distinct functional organization: opponent cells in bottleneck layers, non-opponent cells in subsequent layers.
• Shallower networks without bottlenecks learn complex nonlinear color systems.
• Deeper networks with tight bottlenecks learn simple channel opponent codes.

Conclusions:

• Bottleneck architecture induces strong functional organization in CNNs.
• Methods provide insights into how CNNs encode color and spatial information.
• Findings enhance understanding of the relationship between CNN architecture and learned representations.