Clarfication on Gradient descent for neural networks

There are a lot of threads discussing this. Please check this one, two, three, and four.