A tiny network (2 inputs → 6 hidden → 1 output) trained by gradient descent. Watch the decision boundary bend itself around the data as weights update. Different data shapes need different complexity.
Epoch0
Loss—
Accuracy—
Decision boundary (input space)
Network — signal flowing
6
What hidden units do: each one learns a single half-plane (a linear cut). The output combines them through another linear layer. With ReLU non-linearity in between, you can bend those cuts into circles, spirals, anything. More units = more flexibility = risk of overfitting.