TSMixer

Summary

TSMixer is an all-MLP time-series forecasting architecture that mixes information along time and feature dimensions without self-attention. In this wiki it is mainly important as the compact architectural ancestor of Tiny Time Mixers.

Role In The Wiki

TSMixer is a useful counterweight to Transformer-first assumptions in time-series modeling. It suggests that strong time-step-dependent and feature-mixing priors can be competitive for forecasting, especially when cross-variate and exogenous information are available.

Official Artifacts

• Original paper code: https://github.com/google-research/google-research/tree/master/tsmixer
• PyTorch implementation: https://github.com/ditschuk/pytorch-tsmixer

Evidence

• TSMixer: An All-MLP Architecture for Time Series Forecasting
• Tiny Time Mixers

Related Pages

• Time-Series Foundation Models
• Time-Series Scaling And Efficiency
• High-Dimensional Time Series Forecasting