squeeze-and-excitation networks (senet)

These models introduce a new building block for CNNs that improves channel interdependencies at almost no computational cost. They adaptively recalibrate channel-wise feature responses by explicitly modeling interdependencies between channels.