Runs

Three complete runs on A100 machines, made over the past week. Each viewer replays the fifty recorded decisions of a run from its ledger: the controller's activity on the measured graph, the chosen crystal, the evaluator's answer and the learning update.

RunController graphObjectiveEvaluationsAcceptance
forager_rhea_v116,384 MaleCNS neuronsΔH_mix, 0 K, unrelaxed40 MACE · 10 QE205 checks, all passOpen viewer →
forager_rhea_wholebrain_v1163,972 neurons · 6,143,838 edgesΔH_mix, 0 K, unrelaxed40 MACE · 10 QE205 checks, all passOpen viewer →
forager_rhea_wholebrain_T1500K_P5GPa_v1163,972 neurons · 6,143,838 edgesΔG_mix at 1500 K, 5 GPa40 MACE · 10 QE207 checks, all passOpen viewer →

Reading a viewer

The circuit panel draws a subset of the run's neurons (all neurons with a downloaded skeleton, plus an even sample of the rest) coloured by the magnitude of their simulated rate when the shown candidate was scored, before the evaluator answered and before that step's weight update. Real morphology draws the measured skeletons of 128 neurons from the MaleCNS atlas; Topology is an abstract layout. The selected crystal is the input cell of the chosen candidate. The four figures below the timeline are the evaluator and its evidence label, the acquired objective in eV/atom, the reward-minus-baseline signal, and the cumulative L2 change of the controller's edge weights. The ledger lists every evaluation; click a row to jump to it, or use the arrow keys.

Viewer pages load about 8 MB of recorded data each.