Squeeze-and-Excitation Networks

Convolutional neural networks extract informative features by fusing spatial and channel-wise information within local receptive fields, and while several recent approaches have boosted representational power by enhancing spatial encoding, this paper instead focuses on the channel relationship. The authors propose the 'Squeeze-and-Excitation' (SE) block, a novel architectural unit that adaptively recalibrates channel-wise feature responses by explicitly modeling interdependencies between channels, and show that stacking these blocks together constructs SENet architectures that generalize extremely well across challenging datasets.

Crucially, SE blocks were found to produce significant performance improvements for existing state-of-the-art deep architectures at minimal additional computational cost. SENets formed the foundation of the authors' ILSVRC 2017 classification submission, which won first place and significantly reduced the top-5 error to 2.251%, achieving roughly a 25% relative improvement over the winning entry from 2016.