Depth in convolutional neural networks solves scene segmentation

Area of Science:

• Computer Vision
• Neuroscience
• Artificial Intelligence

Background:

• Deep convolutional neural networks (DCNNs) match human object recognition in natural scenes.
• Human vision requires additional processes for complex scenes beyond feedforward activity.
• The role of network depth in differentiating object from background information is unclear.

Purpose of the Study:

• To investigate how object information is separated from backgrounds in DCNNs of increasing depth.
• To explore the relationship between network depth and scene segmentation capabilities.
• To compare DCNN performance with human visual processing for complex scenes.

Main Methods:

• Controlled object and background information by manipulating noise, congruence, and occlusion.
• Analyzed DCNNs with varying depths to assess object-background differentiation.
• Compared performance of shallow networks trained on segmented versus non-segmented objects.

Main Results:

• Increased network depth led to improved distinction between object and background information.
• Shallow networks benefited from training on segmented objects.
• Deeper networks demonstrated an emergent capability for scene segmentation.

Conclusions:

• Sufficiently deep DCNNs can perform scene segmentation implicitly.
• The human brain may achieve scene segmentation during object identification via feature binding, similar to deep neural networks.
• Explicit segmentation mechanisms may not be necessary for object recognition in complex visual environments.