Machine learning, visualized.

Watch algorithms learn, step by step. Gradient descent rolling, K-means snapping, neural nets bending a decision boundary — every demo derives the math from first principles and shows the algorithm converging iteration by iteration.

8interactive demos
4topics
∞tweaks

Curriculum four chapters, derived from first principles

Foundations

Optimization

How models learn at all. Gradient descent on a loss landscape, parameter updates iteration by iteration.

2 demos~5 min
Unsupervised

Clustering & Reduction

Unsupervised structure. K-means centroids snapping into place; PCA collapsing dimensions while preserving variance.

2 demos~6 min
Supervised · linear

Linear Models

The first classifiers. Perceptron flipping its boundary on every misclassified point; logistic regression sliding into a smooth sigmoid.

3 demos~8 min
Deep

Neural Networks

From perceptron to multi-layer. Visualize fully-connected, LeNet, and AlexNet-style architectures with SVG export.

1 demo~6 min

How to learn here

① From first principles Each algorithm is derived step by step, then implemented in plain numpy — not hidden behind a one-line library call.
② Watch it converge Every iteration re-renders. Pause, step, rewind. The marker, the centroids, the decision boundary all update live.
③ Build intuition Drag a learning rate, click a starting point, edit the loss. Re-run instantly and feel the algorithm respond to your choices.
④ Math + code, side by side The derivation, the source, and the animation on one page. Read the math, then watch it execute on your data.

Frequently asked

Begin with Chapter 1: Gradient Descent. Every other algorithm on this site uses gradient descent (or a relative) to actually learn its parameters, so the intuition you build there carries everywhere. Then walk the chapters in order — clustering, linear models, neural networks.
No. Every demo runs entirely in your browser tab — no install, no Jupyter, no Colab, no signup. Open a page, click Run, watch the algorithm train.
Yes. Every demo page shows the underlying algorithm in an editable panel. Change a learning rate, a kernel, an init seed, or rewrite the loop entirely — hit Run and watch the visualization update with your version.
The demos implement each algorithm with plain numpy (loops, matrix ops, gradient computations) rather than calling a one-line sklearn.fit. The goal is to see the math actually executing — every iteration of gradient descent, every centroid reassignment, every weight update — so you build intuition for what the library calls are hiding.
The curriculum structure is inspired by ml-visualized.com (four chapters: optimization, clustering, linear models, neural networks), but the interactivity is different. ml-visualized embeds pre-rendered animations; here you can pause, edit the algorithm, drag the starting point, and re-run on your own settings.
The existing tools (NN Architecture Visualizer, Activation Function Explorer, Logistic Regression Calculator, ROC/AUC, ML Pipeline) are utility pages that compute a result or render a diagram. The chapter demos are training visualizations — you watch the algorithm iterate and converge, with editable source so you can experiment.