# Is important the role of $b$ to solve SVM in the primal formulation?

Using the Representer theorem we can define a approximation function with or without the bias term $$b$$. If $$b$$ is considered, how should it be computed when training the primal SVM with gradient descent? Does $$b$$ affects the generalization performances?