Attention Mechanisms in CNNs

Attention is intuitively understood from our every day perception experience - we tend to focus our sight into certain parts of the image with high resolution while assigning lower resolution to parts of the image that is not interesting for the task at hand. Such weighting of the input features that we perceive can be mirrored in CNNs with a structure that we call attention map.

We represent attention as a probabilistic map over the input image locations

https://towardsdatascience.com/learn-to-pay-attention-trainable-visual-attention-in-cnns-87e2869f89f1