Hierarchical Latent Structure Learning through Online Inference
Ines Aitsahalia, Kiyohito Iigaya
This paper introduces HOLMES, a computational model that learns hierarchical (multi-level) structures from sequential data in real-time without needing labeled examples. The model combines statistical techniques to discover how experiences relate at different levels of abstraction, helping AI systems generalize knowledge while still distinguishing important details. In tests, HOLMES created more compact representations and transferred knowledge more effectively than simpler flat models.
Bayesian inferencehierarchical learningsequential datalatent variable models