A cooperage trains a network on stave photos paired with stave notes. An apprentice says the run succeeded because the preview “looks nice.” What quantity actually drives the weight updates during training?

During training, the optimizer adjusts parameters to reduce a defined loss on the multimodal batch. A pleasant preview can be unrelated to that objective and does not drive updates. Associating success with the loss—not with a casual visual impression—matches how deep-learning training actually works.