Pooling is a standard part of convnet architecture. Translation invariance is also encoded in parameter sharing: it's the *same* filter you're running across the image (or part of it). Either alone does not do the job.

→ View original post on X — @pmddomingos,