Initialization sets the initial values of model parameters (like weights in a neural network) before training. Good initialization can speed up convergence and improve final performance. Strategies include random uniform, Xavier initialization, or He initialization, depending on the activation function.