Computes gradient moment for a bias variable through an Add operation. Assumes ``Z = tf.add(Zz, b)``, where ``b`` is a bias parameter and ``Zz`` is a ``[n, ?]`` tensor (``n`` being the batch size). Broadcasting for all kinds of shapes of ``Zz`` (e.g. ``[n, d_in]`` or ``[n, h_in, w_in, c_in]
(op, b, out_grad, batch_size, mom=2)
source not stored for this graph (policy: none)