Deeper neural network models better reflect how humans cope with contrast variation in object recognition

Area of Science:

• Cognitive Neuroscience
• Computational Vision
• Visual Perception

Background:

• Human visual perception exhibits remarkable robustness to variations in stimulus contrast, even in challenging conditions like fog.
• Understanding the neural mechanisms underlying contrast invariance is crucial for explaining visual perception.
• The role of different stages within the visual hierarchy in achieving contrast invariance remains incompletely understood.

Purpose of the Study:

• To investigate whether contrast effects diminish along the visual processing hierarchy.
• To determine the contribution of later visual hierarchy stages to contrast invariance.
• To explore the mechanistic basis of contrast invariance using computational models.

Main Methods:

• A behavioral experiment manipulating stimulus contrast and employing backward masking to probe higher-level visual areas.
• Utilizing three state-of-the-art computational models of human vision with varying hierarchical depths.
• Assessing performance in a visual categorization task under different contrast conditions and masking paradigms.

Main Results:

• Backward masking significantly impaired performance, indicating the involvement of higher-level visual areas in contrast invariance.
• Contrast effects were observed across all levels of the visual hierarchy in computational models.
• Deeper hierarchical models, representing later visual stages, demonstrated greater effectiveness in mitigating contrast effects.

Conclusions:

• Contrast effects extend throughout the entire visual hierarchy, from early to later stages.
• Later stages of the visual hierarchy play a critical role in compensating for variations in visual input contrast.
• These findings highlight the adaptive capabilities of the human visual system in maintaining stable perception despite degraded input.