Home▸Machine Learning & Neural Networks▸Two-Weight Neural Network Loss Landscape: SGD, Momentum, RMSprop and Adam Rolling on a 3D Surface

🧠 Two-Weight Neural Network Loss Landscape: SGD, Momentum, RMSprop and Adam Rolling on a 3D Surface

Train a real two-weight neural network four ways at once: SGD, Momentum, RMSprop and Adam marbles roll down the same 3D loss surface. Find the symmetric saddle, the vanishing-gradient plateau and the learning rate that blows up, and check backpropagation against finite differences.

Machine Learning & Neural Networks3DModerate60 FPS📱 Mobile-adapted🟠 Blender

⚙ Under the hood

Neural networkLoss landscapeGradient descentAdamBackpropagation

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install