Кафедра інформатики

Permanent URI for this collection

Now showing 1 - 1 of 1

Massively Parallel MorphoNAS Implementation
(Національний університет "Києво-Могилянська академія", 2025) Medvid, Serhii
Our GPU-accelerated MorphoNAS (Morphogenetic Neural Architecture Search) enhances the biologically motivated developmental model of the original MorphoNAS, from sequential developmental simulations to large scale parallel simulation. Like the original CPU-based MorphoNAS system, the GPU enhanced version models morphogenesis (neural development) through reaction-diffusion, progenitor differentiation and axon-guided wiring that grow recurrent controllers. Unlike the original CPU-based version, the GPU-based version has the ability to formulate growth and control in parallel execution of thousands of genomes. Through use of sparse, device-resident representations of developmental dynamics and recurrent rollouts, we produce behaviorally equivalent recurrent controllers with up to three orders of magnitude speedup. As a result, we have enabled the possibility of performing evolutionary search over populations of developmental programs, transforming MorphoNAS from a proof-of-concept model of artificial morphogenesis and adaptive architecture discovery to a scalable framework for those objectives.