To train the neural network, the researchers relied on gradient descent to minimize the loss function.