Neural network visualizer
Watch a neural network learn - every neuron, weight and gradient, one step at a time, right in your browser.
What you see
A tiny network learns to separate two interleaving half-moons. Pick one point and follow it forward through the network: each weight multiplies, each neuron adds and applies its activation, and the output becomes a prediction.
Then the error flows backward. Every weight shows its gradient and how much it changes - and the decision boundary redraws as the network gets better.
Why it helps
Diagrams of neural networks hide the arithmetic. Here every number is on screen and every calculation is checked, so the idea behind training stops being abstract.
When the small network makes sense, build your own: stack layers, choose activations (even one per neuron), and train on real data - tables, images or text.
Try it in a minute
- Press Try it now - the Two moons data and a small network are set up for you.
- Press Start, then Next step, and follow one point through the network.
- Switch to full-speed training and watch the boundary and the loss curve.
Questions
Is it free?
Yes. Building, training and looking inside networks is free and runs in your own browser - nothing to install.
Do I need to code?
No. Everything is clicks; when you are ready, you can download the whole project as Python code.
Which networks can I visualise?
Dense networks, CNNs, LSTMs / GRUs, Transformers like BERT, and pretrained models such as ResNet and MobileNet.