Watch a neural network learn
This is a real, small feedforward network training in your browser — not a canned animation. Pick a dataset, shape the architecture, and hit play: every weighted sum, every activation, and every gradient you see is the actual number the network is computing on that step. Click any point on the plot to probe it and watch the exact forward pass and, once the network has seen a label, the backward pass — the error signal δ and every weight gradient — flow through the diagram below.
Decision boundary
Background color is the network's live prediction. Dots are training data — click one (or anywhere) to probe it.
Network
Node number = live activation for the probed point. Edge color/thickness = weight sign & size. Click a neuron to inspect it.
Neuron inspector
Training loss
Mean cross-entropy loss over the full dataset, recomputed after every gradient descent step.
What's actually happening
Each neuron computes a weighted sum of its inputs plus a bias — Σ(w·x) + b — then squashes it through an activation function. Stack enough of these and the network can bend a straight decision boundary into whatever shape the data needs. The output neuron uses a sigmoid, so its activation is a probability between 0 and 1.
How training works
The loss (cross-entropy) measures how wrong the prediction was. Backpropagation applies the chain rule layer by layer, starting at the output, to compute exactly how much each weight contributed to that error — the error signal δ. Gradient descent then nudges every weight a small step opposite its gradient, scaled by the learning rate, and repeats.
Reading the diagrams
In the network view, each node's number is its current activation for whatever point you're probing; edge color shows a weight's sign and thickness shows its magnitude. In the decision boundary, color is the model's live prediction across the whole plane — watch it reshape itself as training runs.