#equivariance

Question 1: Single choice

A CNN trained only on upright photos misclassifies the same photos rotated by 90°. Which statement explains this?

Convolution is equivariant to translations but not to rotations, so rotated patterns produce different features1
Max pooling already makes CNNs invariant to rotation, so the error must come from the fully connected classifier head2
CNNs are invariant to every geometric transformation, so the photos must be corrupted3
The rotated images have a different number of pixels, which breaks the convolution4