Home▸AI & Machine Learning▸Gradient Descent and Backpropagation Training Lab

🧮 Gradient Descent and Backpropagation Training Lab

An interactive 3D loss landscape where a marker descends toward a minimum, letting you compare vanilla gradient descent, momentum and Adam on bowl, ravine and multi-minima surfaces.

AI & Machine Learning3DAdvanced60 FPS

A 3D loss landscape built from two weight values, where a glowing marker trains itself downhill step by step so you can watch what backpropagation's gradients and different optimizers actually do to the weights.

🔬 What It Demonstrates

The surface height is the loss for every pair of weights; the marker's exact position is (w1, loss, w2). Every step computes the true gradient of that surface — the same math backpropagation performs — and moves the marker according to the selected optimizer.

🎮 How to Use

Pick an optimizer and a loss surface, then adjust the learning rate and momentum to see their effect live. The ravine surface makes vanilla gradient descent zig-zag, while momentum and Adam glide smoothly toward the floor.

💡 Did You Know?

A learning rate that is too large can make gradient descent overshoot the minimum and diverge entirely — try pushing the learning rate slider high on the bowl surface to watch it happen.

⚙ Under the hood

An interactive 3D loss landscape where a marker descends toward a minimum, letting you compare vanilla gradient descent, momentum and Adam on bowl, ravine and multi-minima surfaces.

neural networksgradient descentbackpropagationoptimizationmachine learningloss functionThree.js

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