Reproduce “Minimal Convolutional RNNs Accelerate Spatiotemporal Learning” 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
Coşku Can Horuz, Sebastian Otte, Martin V. Butz, Matthias Karlbauer
Researchers from the University of Lübeck and the University of Tübingen developed MinConvLSTM, MinConvGRU, and MinConvExpLSTM, a new class of convolutional recurrent neural networks designed to accelerate spatiotemporal learning. These models enable fully parallelizable training by restructuring recurrent gate dependencies, achieving up to a 5x speed-up in training and demonstrating improved or comparable prediction accuracy on Navier-Stokes dynamics and real-world geopotential data.
Key insights
- By modifying recurrent gating mechanisms to be independent of previous hidden states, spatiotemporal ConvRNNs can be made fully parallelizable across the time dimension, significantly accelerating training.
- Simplifying the recurrent structure in MinConvRNNs does not degrade, and can even improve, predictive performance compared to traditional ConvRNNs, particularly in long-term autoregressive forecasting.
- The use of exponential gating (as in MinConvExpLSTM) can further optimize computational efficiency within the log-domain prefix-sum framework for minimal recurrent units.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.