KANLab: A TensorFlow Playground Fork That Shows KAN Training in the Browser

Researchers Lukas Niehaus, Robin Rawiel and Ulf Krumnack have released KANLab, an interactive browser-based tool for training Kolmogorov–Arnold networks (KANs). It is a direct fork of TensorFlow Playground, the sandbox familiar to many from machine learning courses. A small KAN trains in an open tab, and every B-spline activation is redrawn as gradient descent progresses. The project is described in the preprint arXiv 2610.20117 (October 14, 2026), and its author presented the work on Hugging Face Daily Papers. The code is open source under the Apache-2.0 license.

Why KANs are hard to explain in a lecture

KANs build on the Kolmogorov–Arnold theorem: any continuous function of many variables can be assembled from one-dimensional functions and addition. In a classic perceptron, the weights sit on the edges and the nonlinearity is fixed in the nodes. KANs reverse this: the edges carry trainable one-dimensional functions, most often B-splines. The architecture appeared in 2024 and has since accumulated variants and add-ons.

The add-ons are exactly what gets in an instructor's way. Modern KAN implementations include residual connections and trainable scaling parameters. The curve on an edge stops being simply a function of its control points, and it becomes harder for a student to connect the picture to what is actually being optimized.

What the authors removed and why

KANLab deliberately drops both the residual connections and the trainable scaling. The displayed B-spline curve is exactly its trainable control points, with no hidden multipliers. This choice makes the tool simpler than production implementations. But it serves the main goal: letting users see cause and effect rather than an approximation.

The setting is the same as in TensorFlow Playground: a compact two-dimensional classification task, where the entire network fits on screen and can be followed by eye.

What you can do in the interface

The authors list the following features:

• view the activation functions on every edge of the network;
• edit activations by hand, for example by dragging individual control points;
• watch gradient descent correct the manual intervention and return the curve to the optimum;
• choose initialization schemes and compare how they affect the start of training;
• examine activation variance;
• run interactive pruning, that is, remove edges from the network;
• retrain the network after pruning and see what remains.

Pruning and retraining are easy to demonstrate visually in this environment. A student can see which edges carry the load and which have turned into nearly flat curves. Checking how the network recovers after a manually corrupted activation takes about a minute, with no Python and no GPU.

Where to get it and who it's for

The tool is available at kanlab.app, with a demo Space on Hugging Face (luniehaus/kanlab) and a GitHub repository (lniehaus/kanlab). At the time of publication, the repository has about four stars and the paper has one upvote. The project is very young and primarily educational.

The practical value is fairly down-to-earth. For instructors of deep learning courses, KANLab offers a ready-made demo for a lecture or seminar: it runs in any browser, and the Apache-2.0 license allows modifying it and embedding it in course materials.

For researchers just entering the KAN topic, it helps build intuition quickly about the role of control points, initialization and pruning before moving on to libraries with the full set of complications. It is not intended for real-world tasks: two-dimensional toy datasets and a simplified architecture say nothing about how KANs perform on large datasets. The authors are asking for feedback, especially from those who already include KANs in their curricula, and comments can be left in the repository or under the paper on Hugging Face.