Dropout regularization reduces overfitting for two related reasons. First, because a random subset of units is dropped on each training iteration, the network is effectively forced to train using a smaller, thinner sub-network at every step, which limits the capacity available to memorize noise in the training data. Second, dropout tends to shrink the squared norm of the weights, producing a regularizing effect similar to L2 regularization; because no single unit can rely on any other specific unit being present, weight magnitude and influence are spread out across the network rather than concentrated on a few features, which further prevents overfitting.