Changing the architecture rebuilds the weights and resets training.
| x₁ | x₂ | y | ŷ |
|---|
Each training step runs three phases: activations pulse forward (yellow), gradients pulse backward (green, edge glow = gradient size), then weights update via SGD with momentum. Wider layers converge faster; deeper sigmoid networks are harder to train — try 2 hidden layers and watch how much longer it takes than 1.