Neural network training involves adjusting the weights of the network through a process of optimization to minimize the error in predictions on a given dataset. This is typically achieved through iterative algorithms such as gradient descent, which update the weights based on the computed gradients of the loss function with respect to the weights.