Research · Neural DNA
Networks should grow their topology the way brains do.
A tiny learned “genome” decides how a network wires itself: default disconnected, with metabolic cost forcing selectivity. The same idea travels from artificial neural networks to real biological ones. Three papers, each building on the last. Start at Paper 1 if you're new.
354genome parameters
35.4Mconnections wired
99,970:1compression
Paper 1Neural DNA (NDNA)A compact genome for growing neural network architecturePaper 2Scaling Neural DNA to GPT-2354 parameters wire a language model. 99,970:1 compression. Beats GPT-2 on 3 benchmarks.Paper 3NDNA on Protein Interaction NetworksA 290-parameter genome learns which protein interactions matter for cancer, and the same genome works across breast, lung, colon, and prostate cancer.
Papers on Zenodo, code and models on GitHub and Hugging Face. Paper 2's page has an interactive visualization of the topology evolving in real time.