🧪 NAS Performance Estimation: 3D Supernet Board, Weight Sharing vs Early Stopping vs Full Training
A Blender-built supernet board with 6 edges and 4 operations: compare full training, early stopping and weight sharing as estimators of 4096 architectures, with live Kendall rank correlation, an exact main-effect (ANOVA) model of what weight sharing can express, budget-versus-quality curves, and six independent checks.
⚙ Under the hood
A Blender-built supernet board with 6 edges and 4 operations: compare full training, early stopping and weight sharing as estimators of 4096 architectures, with live Kendall rank correlation, an exact main-effect (ANOVA) model of what weight sharing can express, budget-versus-quality curves, and six independent checks.
Neural Architecture SearchPerformance EstimationWeight SharingEarly StoppingKendall TauANOVA Main EffectsSupernetAutoML
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install