Gradient descent is an optimization technique that utilizes the gradients computed from backpropagation to update the weights and biases, moving towards the goal of minimizing the loss. As shown in the following diagram, the cost (or loss) function is minimized by adjusting the weights, along the slope or gradient of the function:

For a simple perceptron, this cost function is linear, with respect to the weights. But for deep neural networks, the cost function is most often high-dimensional and non-linear. As gradient descent has to traverse paths along all of the different dimensions, it may be difficult to arrive at the global minimum in an acceptable time. To avoid this problem and train faster, neural networks normally employ stochastic gradient descent, which is explained next.