Backpropagation is the algorithm used to compute gradients in deep neural networks, enabling efficient training via gradient descent. It works by applying the chain rule of calculus to propagate the error from the output layer back through the network, layer by layer, updating weights to minimize the loss function.

The process involves two main steps: a forward pass to compute predictions and the loss, and a backward pass to calculate how much each parameter contributed to the error. Gradients are then used to update the model weights using an optimizer (e.g., SGD or Adam).

Backpropagation allows deep networks to learn complex functions from data, but it can suffer from issues like vanishing or exploding gradients—especially in very deep or poorly initialized models. Techniques such as ReLU activations, normalization, and residual connections help mitigate these problems.

Despite its conceptual simplicity, backpropagation is a key innovation behind the success of modern deep learning, making it feasible to train large networks end-to-end.