🏁 Optimizer Race 3D: Decision Surfaces of SGD, Momentum and Adam
Three copies of the same small neural network, started from identical weights and fed identical mini-batches, are trained live by SGD, momentum and Adam. Each learns a 3D decision surface over a Blender-modelled plinth, with finite-difference gradient checks and an exact first-step Adam test.
⚙ Under the hood
Three copies of the same small neural network, started from identical weights and fed identical mini-batches, are trained live by SGD, momentum and Adam. Each learns a 3D decision surface over a Blender-modelled plinth, with finite-difference gradient checks and an exact first-step Adam test.
Neural NetworksOptimizersBackpropagationAdamDecision BoundaryThree.js
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install