Meta-Sine Foundry
Meta-Sine Foundry learns an initialization that can specialize to a new sinusoid from five observations and a handful of gradient steps. Tasks vary in amplitude and phase. A first-order MAML learner, an ordinary model trained on pooled tasks, and an untrained architecture-matched control receive the same adaptation rule at test time.
Evaluation covers 200 seeded tasks and reports mean query MSE before adaptation, after one support-set update, and after five updates. The benchmark tests rapid adaptation, not whether the meta-learner has discovered a universal regression prior.
Verified results
Each architecture has 1,761 parameters. Evaluation used 200 unseen tasks, five
support points per task, and the same 0.01 inner learning rate.
| Initialization | 0 updates MSE | 1 update MSE | 5 updates MSE |
|---|---|---|---|
| First-order MAML | 3.1180 | 1.8105 | 0.7417 |
| Pooled pretraining | 3.1389 | 3.6318 | 3.7822 |
| Random initialization | 4.4218 | 4.4110 | 4.6232 |
After five updates, the meta-learned initialization reduced mean query error by 80.39% versus pooled pretraining and 83.96% versus random initialization. Pooled and random controls worsened under the meta-learned step size, which is part of the measured adaptation advantage rather than a claim that they could not be retuned.
Reproduce
uv run python projects/meta-sine-foundry/train.py