ArXiv paper formalizes and generates compositional structure using algorithmic information theory
The paper formalizes compositional structure by defining it as shared structure revealed when parts are modeled together but not separately, using algorithmic information theory to identify such structure across scales. The authors derive tractable approximations and invert the framework to optimize strings and images that exhibit hierarchical part‑reuse, and discuss implications for machine learning including perspectives on network depth and principles for compositional curricula.