That is a concern that is shared with ReLU. But since the weights are shared across the context/minibatch, perhaps that would not be an issue, similar to ReLU.