Does visual experience give rise to geometry?
In this talk, I will discuss two recent projects that aim to analyze the relation between features learned by vision models from 2D observations and the geometry of the 3D world. In the first project we investigate whether vision features learned from the experiences of unique individuals share a common geometric structure and whether this structure enables solving downstream tasks. In the second project, we consider whether it is possible to learn the SE(3) group structure of Euclidean transformations purely from passive visual experience. Both of these projects shed light on whether machine learning models, trained purely through passive observation on massive amounts of data, can "understand" geometry, and whether there are limits to such training.