Harnessing cortical geometry, wiring, and function as inductive biases for recurrent neural networks
Researchers have developed biologically grounded recurrent neural networks by leveraging data from the MICrONS program, which combines electron microscopy and calcium imaging of mouse visual cortex. These networks utilize neuronal spatial coordinates, anatomical connectivity, and function-derived relationships from nearly 12,000 neurons to initialize weights and impose spatial constraints during learning. The study found that networks incorporating cortical structure and function significantly outperformed baseline models across three cognitive decision-making tasks, with functional weight initialization providing the most substantial gains. AI
IMPACT Biologically inspired network architectures may lead to more efficient and effective learning algorithms.