KANLab: TensorFlow Playground for Kolmogorov-Arnold Networks

KANLab. TensorFlow Playground for Kolmogorov-Arnold Networks. By Lukas Niehaus and two others, from the University of Osnabrueck and the University of Applied Sciences Osnabrueck. Published on the fourteenth of October. Its code repository is public and real: thirteen source files.

KANLab is a browser-based visualization tool for teaching, interactively, the core concepts of Kolmogorov-Arnold Networks, known as KAN. It is open-source and available at kanlab dot app.

Everything happens inside a compact two-dimensional machine learning setting. Users can inspect and modify activation functions and initialization schemes, and then watch gradient descent correct the disruptions they introduced by hand. They can also explore activation variance, prune the network interactively, and retrain it.

KANLab is forked from TensorFlow Playground. One design decision sets it apart from modern KAN implementations: it leaves out their residual connections and their learnable scaling parameters. That omission is made for educational clarity, so that the B-spline activation curves shown on screen correspond directly to the trainable control points underneath them.

The tool is built for teaching, so its audience is anyone learning how a KAN works who wants to see the moving parts rather than read about them.

The trade the authors made is fidelity for legibility. Dropping two features of modern implementations means what you see on screen is exactly what is being trained, which is what lets a learner grab a curve, bend it, and see how training answers.