September 06, 2018

SeNet

- arxiv.org/abs/1709.01507;

- A 2017 Imagenet winner;

- Mostly ResNet-152 inspired network;

- Transfers well (ResNet);

- Squeeze and Excitation (SE) block, that adaptively recalibratess channel-wise feature responses by explicitly modelling in- terdependencies between channels;

- Intuitively looks like - convolution meet the attention mechanism;

- SE block:

- pics.spark-in.me/upload/aa50a2559f56faf705ad6639ac973a38.jpg

- Reduction ratio r to be 16 in all experiments;

- Results:

- pics.spark-in.me/upload/db2c98330744a6fd4dab17259d5f9d14.jpg

#deep_learning