In Silico Mapping of Visual Categorical Selectivity Across the Whole Brain

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Understanding functional selectivity in the visual cortex is crucial for brain research.
• Traditional methods rely on predefined hypotheses, limiting discovery.
• Existing data-driven methods often use simplistic linear models.

Purpose of the Study:

• To develop a novel in silico approach for data-driven discovery of category-selectivity hypotheses in the human brain.
• To enable the identification of complex, nonlinear relationships between brain activity and visual features.
• To create a framework for generating and testing new hypotheses about visual processing.

Main Methods:

• Utilized an encoder-decoder transformer model with a cross-attention mechanism for nonlinear mapping.
• Integrated diffusion-based image generative models to synthesize maximally activating stimuli.
• Employed large-scale datasets for in silico validation within and across subjects.

Main Results:

• Revealed novel regions exhibiting complex, compositional selectivity for diverse semantic concepts.
• Demonstrated the model's ability to map high-dimensional network features to brain activity patterns.
• Validated the discovered selectivity hypotheses in silico.

Conclusions:

• The proposed in silico approach provides a powerful, data-driven framework for hypothesis generation in neuroscience.
• This method advances the study of visual selectivity beyond traditional experimental limitations.
• The findings can guide future functional magnetic resonance imaging (fMRI) experiments.