Topo-Omni: Discovering Brain Functional Selective Regions Using a Deep Topological Multimodal Model

Topo-Omni Research Core Guide

The research team proposed the Topo-Omni deep topological multimodal model, which integrates visual, auditory, and language cognitive processing through a single continuous simulated cortex, successfully reproducing human brain functional organization and predicting new brain networks. This model breaks through the limitations of existing unimodal and fragmented topological models, providing a new perspective for interdisciplinary research between neuroscience and artificial intelligence.

Source Information:

• Original author team: arXiv paper authors (arXiv:2606.09770v1)
• Publication time: June 8, 2026
• Original link: http://arxiv.org/abs/2606.09770v1

Research Background and Limitations of Existing Models

Topological Organization Characteristics of the Cerebral Cortex

Adjacent neurons in the cerebral cortex have similar response characteristics, forming spatial organization patterns (such as direction-sensitive gradients in the visual cortex and tone gradients in the auditory cortex), which are key to understanding information processing.

Shortcomings of Existing Topological Models

1. Unimodal limitation: Only handles single inputs (visual/auditory/language), unable to integrate multimodal information (inconsistent with the seamless integration feature of the human brain);
2. Layer-wise independent constraints: Applies spatial constraints to each layer independently, leading to fragmented topological maps and failure to capture the continuity of cortical processing streams and cross-modal integration.

Core Innovations of the Topo-Omni Model

1. Single continuous silicon cortex: Integrates visual, auditory, and language cognitive processing into the same continuous simulated cortex, where each position can respond to multimodal inputs while maintaining smooth spatial transitions;
2. Spatial smoothness objective function: Introduces this constraint by fine-tuning pre-trained models to ensure similar responses from adjacent virtual neurons, spontaneously forming topological organization;
3. Cross-modal shared representation: The same spatial position participates in processing different inputs, with visual/auditory regions being continuous functional gradients rather than isolated modules.

Model Validation and Prediction of New Brain Networks

Reproducing Known Brain Networks

The model's spontaneous clustering is highly consistent with human neuroimaging data:

• Primary sensory regions (visual/auditory cortex);
• Higher cognitive regions (language processing, multimodal integration);
• Hierarchical structure from sensory input to higher cognition.

Causal Intervention Experiments

The effects of activating/inhibiting model clusters are similar to those of human brain interventions (such as transcranial magnetic stimulation):

• Activating clusters biases corresponding perceptual judgments;
• Inhibiting clusters leads to decreased performance in corresponding tasks.

Predicting New Brain Networks

Two new candidate networks were discovered through computational screening and verified by human neuroimaging:

1. Natural Landscape Network: Responds to natural scenes and outdoor environments;
2. Animal Recognition Network: Prioritizes processing visual information of animals.

Theoretical Significance and Cross-Disciplinary Implications

Theoretical Breakthroughs

A single spatial principle can organize representations across modalities and processing stages, supporting the gradient theory (functional differences are continuous gradients rather than separate modules) and challenging the traditional modular theory.

Cross-Disciplinary Reciprocity

• Neuroscience inspires AI: Brain organization principles guide the design of multimodal systems;
• AI feeds back to neuroscience: The model generates testable hypotheses about brain organization, serving as a hypothesis generation platform (e.g., computational experiments predict intervention effects).

Current Limitations and Future Research Directions

Current Limitations

• Simplified model: Does not include details of biological neurons;
• Static structure: Does not capture dynamic processes;
• Limited modalities: Only integrates three modalities (the real brain has more abundant modalities).

Future Directions

• Dynamic expansion: Introduce the time dimension to study changes in topological organization with learning/experience;
• Fine-grained improvement: Increase spatial resolution to capture more detailed functional organization;
• Clinical translation: Explore applications in brain disease understanding and neurorehabilitation.