I study two brains in parallel — the artificial and the biological. On the artificial side, I work on representation geometry and training dynamics in AI; on the biological side, I'm leading data-acquisition technologies for the whole mammalian brain connectomics project. Most of my recent work develops new technologies for studying both.
Selected papers
Science of AI
A label-free phase coordinate β(t)/βc(t) for representation dynamics,
derived from a passive GMM probe and verified across SAEs on Pythia, SSL on
CIFAR-10/100, and grokking on modular arithmetic.
Stop SAE training at 5% on early-to-mid LM layers and recover ≥ 96% per-atom
POS purity at 20× less compute. Five-seed evaluation across Pythia-160M /
410M / 1.4B; degrades with layer depth (99% at L6, 88% at L18).
An activity-dependent rule that prunes weights during training under a fixed
broadcast budget. *Equal contribution.
AI for Science
Recovers a global ordering of tens of thousands of EM sections from sparse
pairwise alignments via graph condensation and super-chains.
VQ-VAE with a transformer prior reaches ~1000× compression on nanoscale EM
volumes while preserving downstream segmentation accuracy.
Other