On the Role of Hierarchies, Abstractions, and Representations of Dynamics in Animal and Machine Learning

Hierarchies, abstractions, and representations of dynamics (HARD) are ubiquitous, both in real-world tasks and in animal neural architectures. This suggests HARD have significant potential to improve sample efficiency, safety, and resilience compared to modern AI trends, which have largely departed from the biologically plausible in favor of architectures that can better leverage massive-scale data and computation. We investigate HARD with a focus on the role of internal models in the form of oscillators, both based on explicit decentralized oscillators and in a fully data driven case.