Why should accuracy be used as loss？

[onepeachbiubiubiu]

It seems 'argmax function' in accuracy calculation cannot backpropagate the gradient,so why accuracy be used as loss?

[Lechatelia]

As you said, the accuracy doesn't backpropagate any gradient; removing it from the training configuration will not affect the training results. In fact, we just calculate training accuracy like a loss calculation and then write its result into the log files and tensorboard as a monitoring metric.