A simple and effective technique that combines the advantage of the attention mechanism and dropout, which can be seamlessly integrated into any convolutional neural network to enable better generalization especially on small dataset, without incurring extra computational burdens.
