The output of a convnet is translation-invariant by design. Individual intermediate layers are equivariant. (Transformers may or may not be invariant/equivariant, btw.)

→ View original post on X — @pmddomingos