Reproduce “Piecewise-constant Neural ODEs” 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
Sam Greydanus, Stefan Lee, Alan Fern
To mitigate the computational expense of general Neural ODEs while retaining continuous-time dynamics, a Piecewise-constant Neural ODE (PC-ODE) model is developed, which achieves 3 to 20 times fewer RNN cell updates and often superior predictive accuracy on complex non-linear time series, such as pixel-based billiards.
Key insights
- Constraining latent dynamics to be piecewise-linear allows for exact and computationally efficient integration, overcoming a major bottleneck of general Neural ODEs while retaining continuous-time benefits.
- A learned, adaptive step size guided by an error tolerance enables dynamic temporal abstraction, where the model efficiently skips predictable periods and focuses computational resources on critical events.
- Complex non-linear observation dynamics can be effectively modeled by PC-ODEs, even with simple piecewise-linear latent dynamics, through the use of non-linear encoders and decoders.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.