Learning systems / fit versus represent

Curve Fitter

A model does not jump to its answer. Scrub through the real updates to see how a curve settles, and why an activation lets the same tiny network bend a boundary.

Watch two hypotheses learn

Both models see the white training dots. Outlined dots stay hidden from fitting and test whether the curve learned something useful.

Current linear and polynomial fits
lineardegree 2trainheld out

Fit progress

epoch 0 of 1,000

startscrub the real updatesbudget

Training loss

linear 0.120degree 2 0.120
Training loss

Each line is a real loss history. The marker follows the scrubber.

Change the hypothesis

Degree adds coefficients. The epoch stays put when you change it, so you can compare how much each hypothesis learned with the same budget.

linear now

y = 0.00 + 0.00x

polynomial now

3 changing coefficients

gradient descent, not a jump to the answer

Linear / held out

0.2043

Degree 2 / held out

0.2043

Every frame is computed in this browser from deterministic data and real batch-gradient updates. The displayed accuracy is a geometry illustration, not a production generalization claim.