A neural network is a stack of simple math units (neurons) whose weights are tuned from examples until its predictions match the data. This free playground trains a tiny network right in your browser: step through one forward pass, loss and backpropagation update at a time, and watch the decision boundary learn to separate two classes.
Pick data and a network shape, then press Step to run one round of training and read what changed.
Filled dots: training data. Hollow dots: test data the network never trains on. Click anywhere to inspect a prediction.
Loss over time: solid = training, dashed = test.
Press Step to run one epoch and see the forward pass, loss, backpropagation and weight update explained.
Pick a dataset (circle, XOR, two blobs or spiral) and how much noise to add.
Choose one or two hidden layers with 1 to 8 neurons each, and a learning rate.
Press Step to run one epoch and read the forward pass, loss, backpropagation and weight update explained.
Press Train to run continuously and watch the decision boundary, loss curve and weights change.
Click the plot to see one prediction neuron by neuron, then copy a share link for your setup.
Things to try: XOR with 1 neuron (it cannot solve it) versus 4; the spiral with two layers of 8; learning rate 3 to see unstable training; lots of noise with 8 neurons to see overfitting.
A neural network is a program that learns patterns from examples. It is built from layers of simple units called neurons; each one multiplies its inputs by weights, adds them up with a bias and passes the result through an activation function. Training adjusts the weights until the network's outputs match the examples.
Backpropagation measures how wrong the network is with a loss function, then works backward from the output layer using the chain rule to find how much each weight contributed to that error. That number is the weight's gradient. Gradient descent then nudges every weight a small step in the direction that lowers the loss.
The learning rate sets how big each weight update is. Too small and training crawls; too large and the loss bounces around or gets worse because each step overshoots. In this playground, try 0.3 for steady learning and 3 to see training become unstable.
Overfitting is when a network memorizes its training examples, including their noise, instead of learning the general pattern. You can spot it when training loss keeps falling while loss on held-out test data stops improving or rises. Smaller networks, more data and less noise help.
Yes. ChatGPT is built on a very large neural network called a transformer, with billions of weights instead of the few dozen here. It is trained with the same basic recipe this playground shows: a forward pass, a loss, backpropagation and gradient descent, repeated over huge amounts of text.
Plain-English definitions of transformers, embeddings and other AI terms.
Open toolTest how well you understand LLMs, machine learning and RAG.
Open toolCompare how much text today's large models can read at once.
Open toolFind the right AI tool for your task in a few questions.
Open toolThe same recipe, scaled up billions of times, powers today's AI breakthroughs. Get the good news in the free daily AI Wins digest.
Get the free digestOnly positive AI stories, delivered daily or weekly. Free, no spam, unsubscribe anytime.