#translation-invariance

Question 1: Single choice

mediumAI-checkedConvolutional Neural Networks

A feature map contains a small circle. You shift the circle by one or two pixels and then apply $2 \times 2$ max pooling (stride 2) several times in a row, to both the original and the shifted map. How do the two final maps compare?

They end up (almost) identical, because repeated pooling discards small shifts1
They keep the shift, just at a lower resolution2
They would come out the same for a shift of any size3
The shifted circle vanishes, since only the original pixels can be maxima4

#translation-invariance

Timed test