LearnDeep learning

Neural network

Stack neurons in layers and watch a network learn shapes that no straight line could ever draw.

Deep learning
Supervised
Classification

Choose the data and press Train. The squares in the panel show what each neuron responds to. Click to add points in the chosen colour.

Live numbers

Accuracy0

Loss0

Steps0

The network. Each square shows what one neuron responds to; lines are weights (pink positive, blue negative).

Controls

Data

Neurons per layer4 Hidden layers1

Activation

Add points as

Try this

Train on the circles until accuracy reaches 95%. The boundary closes into a ring.
Solve XOR (95% accuracy) with only 2 neurons in one layer. If it gets stuck, press "Random start": training can settle in a poor answer.
Pick the spiral and reach 90% accuracy. You will need more neurons, a second layer, or both. If it stalls, try "Random start".

How it works

Simple pieces, complex shapes.

A neural network passes the inputs through layers of neurons. Each neuron draws its own straight line, then bends the result with an activation function such as tanh. The next layer combines those bent pieces into curves, rings and spirals.

Training adjusts every weight to lower the loss, here the cross-entropy between the predictions and the true colours. The gradients come from backpropagation, and the updates use Adam, a refined version of gradient descent. Each square in the panel shows one neuron's output across the whole chart.

More neurons and more layers can fit more complex shapes, but they also need more data and can overfit. The same idea, scaled up to millions or billions of weights, is behind image recognition and language models.