Explicit information for category-orthogonal object properties increases along the ventral stream.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• The ventral visual stream is known to build hierarchical representations for object categorization.
• Previous research has primarily focused on categorization, with less attention paid to how other object properties are represented.
• Understanding the neural basis of representing diverse object properties is crucial for comprehending visual perception.

Purpose of the Study:

• To systematically investigate the capacity of different ventral visual areas to represent category-orthogonal object properties.
• To compare the neural encoding of properties like position, size, and pose across the ventral visual hierarchy.
• To determine if a hierarchical model can explain the observed patterns of information representation.

Main Methods:

• Systematic exploration of multiple ventral visual areas using complex naturalistic stimuli.
• Analysis of population encoding of category-orthogonal object properties (position, size, pose).
• Comparison of neural population activity with human performance patterns.
• Development and testing of a hierarchical neural network model.

Main Results:

• The inferior temporal (IT) cortex population explicitly encodes all measured category-orthogonal object properties, including position, size, and pose, more effectively than earlier ventral stream areas.
• The IT population's representations better predict human performance across these properties.
• A hierarchical neural network model successfully replicates the observed cross-area patterns of information.

Conclusions:

• Behaviorally relevant object properties are extracted in concert throughout the ventral visual hierarchy.
• The IT cortex plays a crucial role in representing a wide range of object attributes.
• A hierarchical computational model provides a plausible explanation for the development of these representations.