Home▸AI & Machine Learning▸Discriminative Fine-Tuning: Layer-Wise Learning Rates

Discriminative Fine-Tuning: Layer-Wise Learning Rates (2D)

A real multi-layer network is fine-tuned live on a two-moons toy dataset with genuine per-layer gradient descent: early layers get small discriminative learning rates, the new head gets the full rate, run side by side against uniform fine-tuning so you can measure the real difference in convergence.

AI & Machine Learning2DAdvanced60 FPS📱 Mobile-adapted⇄ 3D version

This is the 2D counterpart to the 3D discriminative fine-tuning simulator: instead of a stylised layer stack, a real small neural network is trained live, in your browser, with actual backpropagation on a genuine two-moons classification task. Two clones of the same freshly-pretrained network are fine-tuned side by side from identical starting weights — one with ULMFiT-style per-layer discriminative learning rates and a slanted-triangular schedule, the other with one uniform learning rate for every layer — so the loss curves, decision boundaries and final accuracy you see are the real, measured outcome of two different optimisation strategies, not a scripted animation.

⚙ Under the hood

Fine-tune a pretrained network layer by layer: watch a slanted-triangular learning-rate schedule and per-layer discriminative rates flow update pulses from a new task head down through a frozen, general-purpose backbone.

transfer learningfine-tuninglearning rate scheduleULMFiTneural networksmachine learning

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install