Small models.
Clearer thinking.
Make a sample. Train a model. See what changes.
Building / Browser experiment
The fit
Synthetic data · 80 samples · 64 train / 16 test
Training data (64)Test data (16)Model prediction
Training error (MSE)
Test error (MSE)
Steps
Learning trace
MSE · log scale · lower is better
Train TestCurrent model
Experiment controls
Preparing the browser experiment…
Computed in this browser.
No uploads or remote training.
Change one thing.
Watch the evidence.
1SampleChoose a shape and noise level. A seed makes the sample repeatable.
2FitGradient descent adjusts the model using only the 64 training points.
3CompareThe 16 held-out points check fit on unseen data. This is a toy experiment.
Inspect sample predictions
First eight points, shown numerically for the current model. Training and test points are labeled separately.
| x | Observed y | Predicted y | Split |
|---|