Name: Meta-Learning Lab
Author: MySimulator

New task (unseen at meta-training time)

Amplitude: 1.8 Phase shift: 1.2

Both models see only 5 labelled points from this brand-new sine task. The meta-trained model was pre-trained across hundreds of other sine tasks (different amplitude/phase) specifically to find a starting point that adapts fast to any new one — a simplified MAML.

Fit quality

Gradient steps taken0

Random-init loss—

Meta-trained-init loss—

True task (dashed) vs. both models' current fit (5 training points shown)