Deep Residual Network Predicts Cortical Representation and Organization of Visual Features for Rapid Categorization

Area of Science:

• Neuroscience
• Computer Science
• Cognitive Science

Background:

• The brain processes visual information through topographic cortical patterns.
• Understanding how distributed visual representations facilitate object categorization is crucial.

Purpose of the Study:

• To develop and utilize predictive encoding models to map human cortical representations to a large set of visual objects.
• To investigate the organization and semantic relationships within these cortical representations.

Main Methods:

• Established predictive encoding models using a deep residual network.
• Trained models to predict cortical responses to natural movies.
• Mapped human cortical representations to 64,000 visual objects across 80 categories with high throughput.

Main Results:

• Cortical representations spanned both ventral and dorsal pathways, reflecting multiple object feature levels.
• Semantic relationships between object categories were preserved in the representations.
• Object representations in the visual cortex were organized into three main clusters: biological, non-biological, and background scenes.
• Hierarchical clustering within these categories was driven by middle-to-high level object features.

Conclusions:

• The study demonstrates a computational strategy for characterizing cortical organization and visual feature representation.
• This approach enables efficient and accurate mapping of visual objects to brain activity.
• Findings provide insights into the neural basis of rapid object categorization.