Cortex-inspired RNNs beat baselines using real neuronal wiring from 12,000 mouse cells
12,000 neurons from mouse cortex teach AI to learn faster and organize like biology.
Researchers leveraged the MICrONS program's functional connectomics data—spatial coordinates, anatomical connectivity, and functional relationships from ~12,000 co-registered excitatory neurons from mouse visual cortex—to build biologically grounded recurrent neural networks. Imposing communication-aware spatial constraints during learning, these networks consistently outperformed baseline and partially constrained models across three cognitive decision-making tasks. Functional weight initialization provided the largest gain, while real spatial embedding yielded robust additional improvements. The resulting networks developed low-entropy, modular, and small-world organization, and retained strong performance even with restricted positive weight recurrence.
- Functional weight initialization from ~12,000 mouse cortical neurons provided the largest performance gain across all three tasks.
- Real spatial embedding (using actual neuronal coordinates) added robust improvements beyond partially constrained models.
- Networks developed low-entropy, modular, small-world organization and maintained performance with only positive recurrent weights.
Why It Matters
Biological inductive biases could dramatically improve RNN efficiency and interpretability, bridging neuroscience and practical AI design.