Learning Mixtures of Linear Dynamical Systems via Hybrid Tensor-EM Method
Hybrid AI method combines tensor guarantees with EM flexibility to decode brain activity patterns.
Researchers Lulu Gong and Shreya Saxena have introduced a novel hybrid method, 'Tensor-EM,' for learning Mixtures of Linear Dynamical Systems (MoLDS). MoLDS are crucial for modeling complex time-series data, like neural activity, where behaviors switch between different underlying dynamical patterns. The core challenge has been balancing the global identifiability guarantees of tensor-based methods with the fitting flexibility of Expectation-Maximization (EM) algorithms. This new approach directly tackles that by first using constructed moment tensors from input-output data to recover globally consistent initial estimates of system parameters and mixture weights, providing a robust starting point that pure EM methods lack.
The technical innovation lies in this two-stage process: the tensor stage ensures the model is identifiable and avoids poor local minima, while the subsequent Kalman EM stage refines all Linear Dynamical System (LDS) parameters with closed-form updates. Validated on synthetic data, Tensor-EM showed superior reliability and robustness compared to standalone methods. Its real-world impact was demonstrated by analyzing neural recordings from primates performing reaching tasks. The method successfully modeled the data and clustered different movement directions into separate, consistent subsystems, matching the accuracy of supervised fits. This proves MoLDS, learned via Tensor-EM, is a powerful, unsupervised framework for deciphering complex neural computations, with potential applications in neuroscience, robotics, and financial modeling.
- Hybrid method combines tensor-based identifiability guarantees with EM refinement for robust learning.
- Successfully modeled and clustered neural data from primate reaching tasks into interpretable subsystems.
- Outperformed pure tensor or randomly initialized EM methods on synthetic benchmarks for reliability.
Why It Matters
Provides a reliable, unsupervised AI tool for neuroscientists and engineers to model complex, switching behaviors in brain data and other dynamic systems.