Text this: An Efficient Dropout for Robust Deep Neural Networks