Understanding TSMixer- Google AI’s new all-MLP architecture for Time Series Forecasting




Among all the recent developments in AI, there was one publication that was gained some attention by people. In a recent blog post+paper, researchers from Google introduced a new architecture. According to their blog TSMixer: An all-MLP architecture for time series forecasting (and paper with the same name), “Time-Series Mixer (TSMixer), an advanced multivariate model that leverages linear model characteristics and performs well on long-term forecasting benchmarks. To the best of our knowledge, TSMixer is the first multivariate model that performs as well as state-of-the-art univariate models on long-term forecasting benchmarks, where we show that cross-variate information is less beneficial.”

In this article, I will be going over their publications to break down why TSMixer is important, how this architecture is created, and how it holds up against architectures. Given how important Time Series Forecasting is in business use-cases of Machine Learning, you definitely don’t want to miss this. I’m personally interested in this model's ability to take multiple types of information and factor that into inference. If it works, it will enable a new degree of performance in various high-impact fields like health and finance.