Reproduce “Learning to Dissipate Energy in Oscillatory State-Space Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jared Boyer, T. Konstantin Rusch, Daniela Rus
Damped Linear Oscillatory State-Space models (D-LinOSS) from MIT CSAIL enhance previous LinOSS architectures by decoupling latent state energy dissipation from oscillation frequency, enabling the model to express a complete range of stable second-order system dynamics. This improvement yields state-of-the-art results across diverse long-range sequence tasks, such as 68.0% accuracy on UEA time-series classification, while also reducing hyperparameter search space by 50%.
Key insights
- D-LinOSS layers achieve full spectral expressivity, capable of representing the entire complex unit disk of stable eigenvalues.
- Previous LinOSS models are theoretically limited, with their reachable eigenvalues occupying a set of measure zero in the complex plane.
- Decoupling damping provides a more flexible control mechanism over latent state evolution, crucial for modeling complex temporal phenomena.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.