Backpropagation is the algorithm used to compute gradients in neural networks, enabling weight updates to minimize the loss function. It propagates errors backward from the output layer to the input layer. This process is fundamental to the training of modern deep learning models.