The Representational Organization of Static and Dynamic Visual Features in the Human Cortex

Area of Science:

• Neuroscience
• Computer Vision

Background:

• Visual information processing involves static and dynamic properties.
• The visual system's organization for these properties is not fully understood.
• Correlations in naturalistic stimuli complicate separating static and dynamic feature representations.

Purpose of the Study:

• To investigate the neural representation of static and dynamic visual features.
• To utilize deep convolutional neural networks (DCNNs) to disentangle these features.
• To compare DCNN representations with human visual pathway responses.

Main Methods:

• Employed two-stream DCNNs to separate static (frame-based) and dynamic (optic flow-based) features.
• Utilized representational similarity analysis (RSA) to compare DCNNs with neural data.
• Analyzed neural responses from 14 human participants across visual pathways.

Main Results:

• Both static and dynamic visual features are encoded across all visual pathways.
• Distinct visual pathways exhibit overlapping and unique representations of static and dynamic information.
• Ventral and dorsal visual pathways show similar posterior-to-anterior gradients for both feature types.

Conclusions:

• The visual system processes static and dynamic information in a distributed yet differentiated manner.
• Deep learning models can effectively separate and analyze visual features.
• Understanding these representations advances knowledge of visual perception and neural coding.