Toward view-invariant representations of object structure learned using object constancy cues in natural movies

An approach to learning view-invariant object representations was explored based on the learning of legal or naturalistic view transformations in time, learned from the statistical properties of natural movies. A simple cell layer responded to localized oriented image structure, and a complex cell layer learned to respond to those subsets of simple cells with the strongest tendencies to trade off activity with each other in response to movement of the visual stimulus. Tradeoffs between simple cells were strongest in response to same-orientation translation, and fell off rapidly with changes in orientation. The local complex cell responses thus became insensitive to typical object motion, evidenced by broadening of response to stimulus phase, while remaining sensitive to local object form. The model makes predictions about synaptic learning rules in complex cells, and mechanisms of successive view-invariance in the primate ventral stream.