Harnessing cortical geometry, wiring, and function as inductive biases for recurrent neural networks

This study presents recurrent neural networks (RNNs) that leverage cortical geometry, wiring, and functional organization derived from the MICrONS program's functional connectomics data of mouse visual cortex. By initializing recurrent weights based on anatomical connectivity and imposing spatial constraints during learning, the biologically grounded RNNs significantly outperformed baseline models across cognitive tasks, demonstrating improved learning efficiency and convergence towards biological computation principles. This approach highlights the potential of integrating biological insights into AI model design, enhancing performance and organizational structure in neural network architectures.