Generalizing biological surround suppression based on center surround similarity via deep neural network models

Area of Science:

• Neuroscience
• Computational Vision
• Artificial Intelligence

Background:

• Context significantly influences sensory perception, particularly in vision.
• Primary Visual Cortex (V1) models often use divisive normalization for surround effects.
• Surround effects in higher visual areas beyond V1 remain less understood, especially for complex stimuli.

Purpose of the Study:

• To investigate contextual neural surround effects in higher visual areas using deep neural networks.
• To generalize understanding of surround effects beyond V1 with complex stimuli.
• To develop and utilize a visualization technique for analyzing neural network responses to surround stimuli.

Main Methods:

• Utilized feedforward deep convolutional neural networks (CNNs).
• Developed a gradient-based technique to visualize excitatory and suppressive surrounds.
• Analyzed network responses to varying center and surround stimuli.

Main Results:

• CNNs replicated key surround effects seen in V1, distinguishing salient center stimuli.
• Networks showed suppression when center and surround stimuli were similar.
• In deeper layers, suppressive surrounds dynamically adapted to changes in center stimuli.
• Untrained networks did not exhibit center-surround similarity-based suppression.

Conclusions:

• Deep neural networks offer a valuable framework for understanding surround effects in complex visual processing.
• The findings provide testable hypotheses for surround mechanisms in higher visual cortices.
• The visualization approach can aid future biological experimental designs for studying visual context.