Here the lower level layers are frozen and are not trained, only the new classification head will update itself to learn from the features provided from the pre-trained chopped up model on the left.