Generalization of data-driven dynamics-learning methods to large-scale biological systems

Determine whether data-driven techniques for learning system dynamics—specifically approaches such as Hidden Markov Models, Switching Linear Dynamical Systems, and recurrent neural network–based methods including reservoir computing—can generalize to large-scale problems in complex biological systems.

Background

The paper discusses the challenge that, unlike in physics where laws of motion are often explicitly specified, many biological systems lack known governing equations. As a workaround, the author reviews data-driven methods that learn a system’s dynamics from observations, including Hidden Markov Models, Switching Linear Dynamical Systems, and recurrent neural network–based approaches (e.g., reservoir computing).

Despite progress, the author highlights uncertainty about the scalability of these methods to large, complex systems, noting potential issues such as identifiability and statistical power. This raises an explicit open question about their ability to generalize to large-scale biological problems.