Why biases are excluded from l2_regularization term?

[Alireza Khorshidi]

Why l2_regularization = tf.reduce_sum(tf.square(W_fc)) and not l2_regularization =tf.reduce_sum(tf.square(W_fc))+tf.reduce_sum(tf.square(b_fc)) here?

[Zachary Ulissi]

My sense from the literature / tutorials / etc was that biases are usually excluded from regularization.

https://stats.stackexchange.com/questions/153605/no-regularisation-term-for-bias-unit-in-neural-network

On a related note, I'm planning to add another flag for type of regularization (L1 or L2)

[Alireza Khorshidi]

Okey, but it seems to be added in all layers other than input, like here or here.