Layers of artificial neurons — each one computes a weighted sum, applies an activation, and passes the signal forward. Watch it happen in 3D.
A neural network is a function — stacked layers of simple computations that, together, can approximate almost any mapping. Each neuron multiplies its inputs by learnable weights, adds a bias, passes the result through an activation function, and hands it to the next layer. During a forward pass, information flows left to right. During training, backpropagation flows gradients right to left.
A 3 → 4 → 3 → 2 network. Press Play to watch the signal propagate layer by layer. Step manually to inspect each transition.
Every network — from a simple perceptron to GPT — is assembled from these four ideas.
Takes several inputs, multiplies each by a learned weight, adds a bias, then squashes the result through a non-linear function. One neuron is like one logistic regression.
Without activation functions, stacking layers is equivalent to one linear layer — no matter how deep. Activations introduce the non-linearity that lets networks learn complex shapes.
Input layer receives raw features. Hidden layers learn intermediate representations. Output layer produces predictions. Depth lets the network learn hierarchical abstractions.
After a forward pass, the error is computed. Backprop uses the chain rule to send gradients backwards through each layer, so every weight knows how much to change.