36 min
Seminar, Hosted by Simons Institute for the Theory of Computing
From "Umwelt" to "World" models
Google DeepMind
Recording
Abstract
Learning objects without supervision remains difficult outside restricted settings despite their central role in human vision. This talk argues that assumptions built into many object-centred models conflict with real-world data, making the usual problem formulation inadequate at scale. It examines that mismatch and presents ongoing work intended to support more meaningful learned representations.