#squeeze-and-excitation

Question 1: Single choice

A squeeze-and-excitation (SE) block global-average-pools each channel of a feature map, passes the pooled vector through two fully connected layers, and rescales channel $c$ by a weight $s_c = \sigma(\cdot)_c$ given by a sigmoid. Read as an attention mechanism, how do these weights differ from softmax attention pooling?

They are sigmoid gates, one per channel, that need not sum to 1, so several channels can stand out1
They form a softmax over spatial positions, like attention over the patches of an image2
They are fixed after training and do not depend on the input image3
They pick exactly one channel, a hard form of attention that sets every other channel to exactly zero4

#squeeze-and-excitation

Timed test