Deep Learning for Computer Vision

Attending regions for classification

An attention mechanism can be used for classification. Attention mechanisms replicate the human behaviour of focusing on regions for recognition activities. Attention mechanisms give more weight to certain regions than others. The method of weight is learned from the data while training. Attention mechanisms are mainly of two types, namely:

• Soft attention: Deterministic in character, this can hence be learned by back-propagation.
• Hard attention: Stochastic in character, this needs complicated mechanisms to learn. It is also expensive because of the requirement of sampling data. 

Following is a visualization of soft attention:

Reproduced from Sharma et al.

The CNN features are computed and weighted according ...