Том 8

Permanent URI for this collection

https://ekmair.ukma.edu.ua/handle/123456789/38090

Browse

Browsing Том 8 by Author "Medvid, Serhii"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    MorphoNAS-Bench: a Benchmark Suite for Morphogenetic Neural Network Generation
    (2025) Medvid, Serhii
    We present MorphoNAS-Bench, a benchmark and toolkit for neural architecture search (NAS) using a generative, developmentally inspired design space. Unlike current NAS benchmark datasets (NAS-Bench-101, NATS-Bench) that use static graph encodings of networks, in MorphoNAS-Bench networks are simple, compact genomes that drive morphogenetic development, allowing for a variety of richly defined, spatially embedded recurrent architectures that emerge through different forms of deterministic growth. The following local developmental rules are used in MorphoNAS to grow genomes: morphogen diffusion, cell division, differentiation, and axon guidance as key mechanisms. The seed benchmark dataset presented in this work consists of 1,000 genome-architecture pairs, taken from a pool of over 50,000 generation attempts using the following quality thresholds: a minimum 5 neurons, 3 edges, and 70% out-degree coverage. The dataset was constructed using Latin Hypercube Sampling (LHS) with orthogonal array design to ensure comprehensive parameter space coverage. The attempts were conducted using both fully stratified parameter sampling and a biologically inspired Genome.random() sampling method, ensuring a reasonable level of coverage of the search space while being plausible. Each sample includes detailed annotations of graph entropy, hierarchy scores, core-periphery structure, transitivity, reciprocity, and structural balance metrics. We share an analysis of the emergent properties like size, modularity, grouping, and efficiency, demonstrating that both generation strategies can produce structured networks that are rich in their nontriviality. The provided Python toolkit provides the means of investigation to test how genomes develop into neural networks, with associated structural analysis, framing MorphoNAS-Bench as a reproducible and biologically inspired testbed for any research studies exploring architecture diversity, evolution, and emergent structure in NAS.