It’s a method of evaluating how machine learning techniques work on the training data. If predictions are good loss function will be low and vice-versa. Gradually, with the help of some optimization function, loss function learns to reduce the error in prediction.