Occam's Razor for Self Supervised Learning: What is Sufficient to Learn Good Representations?