Predicting Identity-Preserving Object Transformations in Human Posterior Parietal Cortex and Convolutional Neural...

Area of Science:

• Neuroscience
• Computer Vision
• Cognitive Science

Background:

• Previous studies established linear mapping of visual object responses in human occipito-temporal cortex (OTC) across non-identity feature changes.
• These changes include Euclidean (position, size) and non-Euclidean (image statistics, spatial frequency) features, indicating near-orthogonal representations.

Purpose of the Study:

• To extend previous findings by examining feature mapping in the human posterior parietal cortex (PPC).
• To investigate responses in pre-trained convolutional neural networks (CNNs) as models of the primate ventral visual system.
• To compare representations across OTC, PPC, and CNNs.

Main Methods:

• Applied general linear mapping functions to link visual object responses across nonidentity transformations in human PPC and five CNNs.
• Analyzed both Euclidean and non-Euclidean feature changes.
• Compared results from PPC and CNNs with those from OTC.

Main Results:

• A linear mapping function successfully linked object responses in PPC and CNNs across Euclidean and non-Euclidean feature changes.
• Object identity and non-identity features are represented in a near-orthogonal manner in PPC and CNNs, mirroring OTC findings.
• Identified specific similarities and differences in visual object representation among OTC, PPC, and CNNs.

Conclusions:

• Human PPC and CNNs exhibit similar representational structures for visual objects as OTC, with near-orthogonal feature encoding.
• These findings highlight conserved and distinct mechanisms for visual object processing across different brain regions and computational models.