DYSCO by Muratore and Weygandt extracts governing equations from dynamics
New algorithm accurately recovers latent trajectories from noisy data.
In their recent paper, Paolo Muratore and Mackenzie Weygandt unveil DYSCO, an innovative multi-view temporal contrastive learning algorithm designed to extract governing equations from latent dynamics. This method addresses the challenge of identifying latent dynamical systems from high-dimensional, noisy measurements by utilizing multiple independent views of the same process. The algorithm effectively disentangles signal from noise, enabling users to recover latent trajectories and governing dynamics with strong theoretical guarantees, even in the presence of nonlinear observations.
DYSCO's performance is validated through empirical testing across a range of dynamical regimes, including chaotic, oscillatory, and metastable systems, under both Gaussian and Poisson noise conditions. This adaptability is particularly beneficial for applications in neuroscience, where Poisson noise is prevalent in neural recordings. By parameterizing dynamics within a structured functional basis, DYSCO facilitates symbolic recovery of governing equations, making it a significant advancement in the fields of representation learning, system identification, and scientific discovery.
- DYSCO leverages multi-view contrastive learning for robust identification of latent dynamics.
- The algorithm demonstrates strong identification guarantees even under noisy, nonlinear conditions.
- Empirically validated across chaotic and oscillatory regimes, enhancing applications in neuroscience.
Why It Matters
This advancement enables more accurate modeling of complex systems, benefiting scientific research and data analysis.