Normalization scales input data to a specific range or distribution, such as [0,1] or a zero mean with unit variance. It helps stabilize training and can improve convergence speed for many algorithms. Common methods include min-max scaling and standardization (z-score).