Reproduce “Deep networks for system identification: a Survey” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Gianluigi Pillonetto, Aleksandr Aravkin, Daniel Gedon, Lennart Ljung, Antônio H. Ribeiro, Thomas B. Schön

A comprehensive survey details the application of deep learning in system identification, outlining various network architectures, optimization methods, and theoretical implications such as overparameterization for modeling dynamic systems from data. It bridges traditional system identification with modern deep learning practices, clarifying benefits and challenges in capturing complex system dynamics.

Key insights

  • Deep neural networks offer a significantly richer hypothesis space for black-box modeling of dynamic systems, capable of capturing complex nonlinearities and temporal dependencies often intractable for traditional methods.
  • Overparameterized deep learning models can achieve strong generalization performance, even when perfectly interpolating noisy training data, due to implicit regularization properties of optimization algorithms and the double-descent phenomenon.
  • Various deep architectures (e.g., CNNs for temporal patterns, RNNs for sequential dependencies, VAEs for latent states) can be effectively adapted to specific aspects of dynamic systems, expanding the range of identifiable models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.