Reproduce “Multi-Time Attention Networks for Irregularly Sampled Time Series” 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
Satya Narayan Shukla, Benjamin M. Marlin
Multi-Time Attention Networks (mTANs) introduce a deep learning framework for multivariate, sparse, and irregularly sampled time series using learnable continuous-time embeddings and attention mechanisms. This approach achieves superior interpolation and classification performance, notably an AUC of 0.858 on PhysioNet mortality prediction, while reducing training times by one to two orders of magnitude compared to ODE-based models.
Key insights
- Learning continuous-time embeddings and attention-based similarity kernels directly from data allows for more flexible and robust modeling of irregular time series than fixed kernels or positional encodings.
- A VAE-based encoder-decoder architecture with a temporally distributed latent representation effectively captures local structure and fine-grained temporal details in sparse data.
- mTANs offer a computationally efficient alternative to ODE-based models, achieving comparable or superior performance with significantly reduced training times through their attention mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.