🧬 3D Differentiable Architecture Search: Bilevel Alpha Gradient, Mixed Edges and the Discretisation Gap
Four marble nodes joined by bundles of four candidate-operation tubes, built in Blender: train a toy DARTS supernet with first- or second-order architecture gradients, watch each softmax bundle collapse, and measure the gap between the mixed network and the discrete cell it is turned into, with brute-force gradient checks.
⚙ Under the hood
Four marble nodes joined by bundles of four candidate-operation tubes, built in Blender: train a toy DARTS supernet with first- or second-order architecture gradients, watch each softmax bundle collapse, and measure the gap between the mixed network and the discrete cell it is turned into, with brute-force gradient checks.
Differentiable Architecture SearchDARTSNeural Architecture SearchBilevel OptimisationSoftmax RelaxationSecond-Order GradientDiscretisation GapAutoML
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install