generalopinionbullishAligning representations at the object level rather than whole scenes promises greater data efficiency in self-supervised learningMachine Learning27 Jul 2026http://arxiv.org/abs/2607.02404v1