Deep learning: build a network and train it
Add layers, add neurons, pick an activation, and train. The point is not to build something useful — it is to develop intuition for what each control does, because in a real project you will be choosing these blind. Watch a network with too little capacity fail to separate the classes at all, then watch one with too much learn the training points perfectly and the boundary badly. Notice how much the learning rate matters compared with the architecture. Most practical deep-learning work is this loop at a larger scale, and the judgement it takes is far cheaper to build here, where a run finishes in a second.
About this tool
What it does
Build a small neural network — layers, neurons, activation — and watch it learn a classification task live, with the network diagram, decision boundary and loss curve updating as it trains.
Who it is for
Anyone who wants to see what “training a model” means. Curiosity is the only prerequisite; the maths is on screen, not required going in.
Example
You enter:Two hidden layers, 6 neurons each, ReLU — then press train.
You get:The decision boundary bending to separate the two classes while the loss curve falls. Cut it to one neuron and watch it fail — which is the lesson.
What it will not do
A tiny network on a 2-D toy problem, trained in your browser in seconds. Real models are orders of magnitude larger and the hard parts — data, evaluation, deployment — are not here. It teaches the mechanism, not the job.
Your data
Training runs entirely in your browser. Nothing is sent anywhere or stored.
Next step
Deep Learning
Build & train a neural network
Stack layers of neurons, choose an activation, and press Train. Watch the decision boundary bend to separate the classes while the loss falls — that's a network learning, live.
More neurons and layers can carve more complex boundaries — but train longer and risk over-fitting.