On learning what to learn: Heterogeneous observations of dynamics and establishing possibly causal relations among them
2024

Learning What to Learn: Understanding Dynamics and Causality in Observations

publication Evidence: moderate

Author Information

Author(s): David W. Sroczynski, Felix Dietrich, Eleni Koronaki, Ronen Talmon, Ronald R. Coifman, Erik Bollt, Ioannis G. Kevrekidis

Primary Institution: Princeton University

Hypothesis

How can we identify the relevant inputs and outputs for a function in a data-driven manner using heterogeneous observations?

Conclusion

The study demonstrates methods for identifying common observables from heterogeneous data streams and learning the relationships between them.

Supporting Evidence

  • The study identifies common observables from different sensors measuring the same phenomenon.
  • It demonstrates the ability to learn functional relationships between measurements from different sensors.
  • The methods can potentially uncover causal relationships in the data.

Takeaway

This study shows how to figure out what information is important from different measurements of the same thing, helping us understand how they relate to each other.

Methodology

The study uses data-driven techniques to process simultaneous heterogeneous data streams and learn input-output relations using various methods like k-nearest neighbors and neural networks.

Limitations

The approach may struggle with 'dirty' channels where measurements are mixed and not easily separable into common and uncommon parts.

Digital Object Identifier (DOI)

10.1093/pnasnexus/pgae494

Want to read the original?

Access the complete publication on the publisher's website

View Original Publication