Hi Haoyu,
the loss is just a function that is getting minimized during training. There is notthing special in the value 0.
Typically, you can define it in a way that makes 0 mean "no error", "perfect result" or a similar concept but it's not necessary for the training to converge.
