step 0
τ = 1.00
τ = 1.00
Cell node
Gradient particle
⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.
This simulation investigates the DARTS technique for automatically designing neural network architectures: a small cell graph — two input nodes and two intermediate nodes joined by five candidate edges — starts with every edge holding a soft mixture over four operations (none, skip-connection, 3×3 convolution, 5×5 convolution). Each training step runs one analytic gradient-descent update on the mixture weights, pulled toward whichever operations best trade off accuracy against compute cost under your chosen efficiency weight, while a decaying softmax temperature sharpens the blend into a single discrete architecture — the exact continuous-relaxation trick real automated machine learning pipelines use to search architecture space with ordinary backpropagation instead of training thousands of candidates from scratch.