Reproduce “Efficient Time Series Processing for Transformers and State-Space Models through Token Merging” 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
Leon Götz, Marcel Kollovieh, Stephan Günnemann, Leo Schwinn
Introducing local and causal token merging techniques significantly enhances the efficiency of Transformer and State-Space Models for time series data. These methods achieved up to 54.76 times faster inference on foundation models like Chronos and maintained or improved predictive accuracy across various benchmarks, including up to 3.80 times for other Transformers and 3.62 times for State-Space Models.
Key insights
- Local merging can enhance forecasting quality by functioning as an adaptive low-pass filter, smoothing time series by combining similar tokens and reducing noise.
- Datasets exhibiting higher spectral entropy and Total Harmonic Distortion (THD) show greater accuracy improvements from local merging, indicating its ability to remove redundant information from complex signals.
- The efficiency gains and improved scaling to longer sequences offered by token merging are superior to merely adjusting the input sequence length in conventional models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.