Backpropagation of an error. Why such a difference in the signs of the formula with the same approach?

Question

Machine learning

I welcome everyone.
When studying neural networks and storming formulas on different sources, an ambiguous question arose with the backpropagation method. For example:
d - desired result y - network response w - weight g - gradient v - sigmoid value
somewhere they write that (1)[e = d - y], and somewhere (2)[y - d].
In weight adjustment (1)[w = w - learning rate * g * v] , and
where then (2) [w = w + learning rate * g * v]; So far, I have only found a regularity in the sign in adjusting the weights with the order of the variables in finding the error.

Answer 1

Because this is not a strict formula, ala "how to find the subtrahend". Here it is important only in which direction it is necessary to "push" the network. And now how much everyone decides for himself in his particular case. And you are not yet using regulation of the first and second levels.