In developing brains, axonal projections follow chemical gradients shaped by local interactions. This paper asks whether such a process can be inferred from its outcome: For instance, given observed mouse brain connectivity, can one recover the developmental program that produced it? If such developmental programs can be recovered, they not only explain how biological connectivity arises, but also offer a biologically grounded search space for artificial intelligence, in which architectures emerge through the evolution of genetic encodings that produce plausible wiring diagrams. A framework is proposed that uses the biological connectome itself as a beacon to guide this search, referred to as the Connectome-Generating, AI-Generating Algorithm (CONGA). Implemented as neural cellular automata (NCA), a model was trained to reproduce axon-tracing data in the mouse connectome, and its internal representations were compared to gene expression patterns measured in the same spatial coordinates. The result demonstrates how the brain of an intelligent organism may self-assemble through an indirect encoding of connectivity. The model outperformed a static linear baseline, but only when constrained in size, suggesting that compact developmental programs better align with biological mechanisms.

PDF

In Proceedings of the Genetic and Evolutionary Computation Conference Companion: Workshop on Evolution and Self-organization, 2025.

@inproceedings{warner:geccows25, title={Self-Organizing Models of Brain Wiring: Developmental Programs for Evolving Intelligence}, author={Jamieson Warner and Risto Miikkulainen}, booktitle={Proceedings of the Genetic and Evolutionary Computation Conference Companion: Workshop on Evolution and Self-organization}, month={ }, url="http://nn.cs.utexas.edu/?warner:geccows25", year={2025} }