Neural DNA: A Compact Genome for Growing Network Architecture

We introduce Neural DNA (NDNA), a compact learned genome of fewer than 300 parameters that grows neural network topology through type-based compatibility rules, default-disconnected initialization, and metabolic cost pressure. NDNA consistently outperforms random sparsity by 0.39% to 7.01% across three architectures (MLP, CNN, Transformer) and five datasets (MNIST, CIFAR-10, CIFAR-100, Fashion-MNIST, IMDB). Topology transfers across tasks without modification. Compression ratios scale with network size, reaching 8,384:1 on the largest architecture tested.