Sergii Medvid

Mykola Glybovets, Sergii Medvid

While biological neural networks develop from compact genomes using relatively simple rules, modern artificial neural architecture search methods mostly involve explicit and routine manual work. In this paper, we introduce MorphoNAS (Morphogenetic Neural Architecture Search), a system able to deterministically grow neural networks through morphogenetic self-organization inspired by the Free Energy Principle, reaction-diffusion systems, and gene regulatory networks. In MorphoNAS, simple genomes encode just morphogens dynamics and threshold-based rules of cellular development. Nevertheless, this leads to self-organization of a single progenitor cell into complex neural networks, while the entire process is built on local chemical interactions. Our evolutionary experiments focused on two different domains: structural targeting, in which MorphoNAS system was able to find fully successful genomes able to generate predefined random graph configurations (8-31 nodes); and functional performance on the CartPole control task achieving low complexity 6-7 neuron solutions when target network size minimization evolutionary pressure was applied. The evolutionary process successfully balanced between quality of of the final solutions and neural architecture search effectiveness. Overall, our findings suggest that the proposed MorphoNAS method is able to grow complex specific neural architectures, using simple developmental rules, which suggests a feasible biological route to adaptive and efficient neural architecture search.

Sergii Medvid, Andrii Valenia, Mykola Glybovets

Developmental approaches to neural architecture search grow functional networks from compact genomes through self-organisation, but the resulting networks operate with fixed post-growth weights. We characterise Hebbian and anti-Hebbian plasticity across 50,000 morphogenetically grown recurrent controllers (5M+ configurations on CartPole and Acrobot), then test whether co-evolutionary experiments -- where plasticity parameters are encoded in the genome and evolved alongside the developmental architecture -- recover these patterns independently. Our characterisation reveals that (1) anti-Hebbian plasticity significantly outperforms Hebbian for competent networks (Cohen's d = 0.53-0.64), (2) regret (fraction of oracle improvement lost under the best fixed setting) reaches 52-100%, and (3) plasticity's role shifts from fine-tuning to genuine adaptation under non-stationarity. Co-evolution independently discovers these patterns: on CartPole, 70% of runs evolve anti-Hebbian plasticity (p = 0.043); on Acrobot, evolution finds near-zero eta with mixed signs -- exactly matching the characterisation. A random-RNN control shows that anti-Hebbian dominance is generic to small recurrent networks, but the degree of topology-dependence is developmental-specific: regret is 2-6x higher for morphogenetically grown networks than for random graphs with matched topology statistics.