AI Now Designs Better AI by Growing Its Own Family Tree
Better blueprints could mean faster, cheaper medical imaging AI — built with less human effort.
WHAT HAPPENED: Researchers introduce EvoTreeNAD, a genealogy-guided evolutionary algorithm that builds trainable neural architectures without a supplied seed or hand-specified search space. Starting from an empty root, it grows a persistent genealogy in which each node is a complete architecture. Top-percentile values computed from each node and its descendants guide lineage selection; an Idea Agent proposes a variant from that history and a Code Agent implements it. Each evaluated variant becomes a child node.
WHY YOU CARE: EvoTreeNAD discovered architectures that outperform the compared NAS and NAD baselines, reaching CIFAR-10/100 test errors of 2.05±0.06% and 15.09±0.22%, and surpassing the strongest listed baselines on all six MedMNIST-v2 tasks. A controlled CIFAR-10 study shows it outperforms direct generation, best-of-N greedy continuation, and full-family-mean routing.
THE CATCH: Evaluation is costly, which constrains the scope of exploration, and the theoretical analysis establishing stationary variation regimes holds only under specified variation assumptions.
- A new system lets AI design other AI, without a human giving it a starting design or a list of options to pick from.
- Its designs reached 2.05% error on a standard image test and beat all listed rivals on six medical-imaging tasks.
- It works like a family tree: designs that perform well get more chances to 'have children,' so good ideas build on themselves.
Why It Matters
Could speed up cheaper, more accurate medical-imaging AI — trimming the expert time needed to design it.