Learning Mixtures of Linear Dynamical Systems

A project on unsupervised clustering of trajectories drawn from linear dynamical systems

Final project for EECS 553 (theoretical graduate-level machine learning).

Contributions and achievements: Chose the problem, explained the theoretical intuition behind the algorithm to teammates, wrote about a third of the code. Found important discrepancies between theory and experiments that were not reported in the paper, corresponded with the authors. Added theory-informed heuristics to determine hyperparameters. Performed ablation studies to compare performance to merely using random-subspace-based dimensionality reduction. Repository link: https://github.com/Chinmaya-Kausik/learning_mixtures_lds_py

Positives: Very fun collaboration, was able to drive progress since it was directly related to my research.

Negatives: Did not get to build experience in a new area of machine learning.