Quick Summary:

 

 

 

Detailed Derivation:

 

 

Even if you look at gradient descent below, error is multiplied by previous value. When input is higher it’s contribution to error is higher and will needs to change more.