Why do we need a bias parameter in neural networks?
Changing the weight of the function changes the steepness of the sigmoid, but sometimes we need to be able to shift the function, right wards or leftwards. For this we need the bias parameter.
Problem visualised without a bias parameter from a Stack Overflow answer:
Another answer.
Note: Instead of intending to write long posts and not writing anything, I will now intend to write more “Bangladeshi train station” like posts ala Tyler Cowen.