spline-kan-pocket-lab

Spline KAN Pocket

Spline KAN Pocket compares a learnable edge-spline network with an exactly parameter-matched tanh MLP on nonlinear symbolic regression. Both models contain 641 trainable parameters and receive the same number of optimization steps.

The benchmark separates interpolation inside the training square from extrapolation in the surrounding ring. The Space renders the true and learned surfaces side by side.

This is a compact piecewise-linear spline KAN experiment, not a claim of faithfully reproducing every implementation detail of a particular KAN library.

Verified local result

At exactly 641 parameters per model, the spline KAN reached 0.00750 interpolation RMSE versus 0.0192 for the MLP. Outside the training grid, however, its RMSE rose to 0.991 versus 0.606 for the MLP, exposing the spline basis's extrapolation boundary.

uv run python projects/spline-kan-pocket/train.py
uv run pytest tests/test_spline_kan_pocket.py

Hosted showcase

This free static Space preserves the complete original Gradio source, trained artifacts, evaluation files, and local launch requirements. Hugging Face now requires PRO for CPU-backed Gradio hosting, so the public landing page is static while the checked-in app.py remains the authoritative runnable demo.