TiDHy: timescale demixing via hypernetworks to learn simultaneous dynamics from mixed observations

Area of Science:

• Computational neuroscience
• Machine learning
• Dynamical systems analysis

Background:

• Neural activity and behavior involve multiple concurrent systems with varying timescales.
• Current methods often oversimplify these dynamics to a single timescale.
• Analyzing complex spatiotemporal data requires methods that can disentangle simultaneous processes.

Purpose of the Study:

• To introduce timescale demixing via hypernetworks (TiDHy), a new computational method.
• To decompose spatiotemporal data into multiple latent dynamical systems with different timescales.
• To accurately model and analyze complex neural and behavioral data.

Main Methods:

• Training a hypernetwork to dynamically reweigh linear combinations of latent dynamics.
• Applying TiDHy to synthetic data with multiple independent switching linear dynamical systems.
• Validating TiDHy on simulated locomotion and real multi-animal social behavior datasets.

Main Results:

• TiDHy accurately reconstructs data and converges to true latent dynamics.
• The method successfully demixes dynamics across orders of magnitude in timescales.
• TiDHy captures fast movement kinematics and slow terrain dynamics in simulated locomotion.
• Extracted dynamics from social behavior data accurately identify mouse social interactions.

Conclusions:

• TiDHy is a powerful algorithm for demixing simultaneous latent dynamical systems.
• The method offers accurate data reconstruction and captures multiple timescales.
• TiDHy has broad applications in computational neuroscience and other domains analyzing spatiotemporal data.