OTIS: Tiny Encoder Achieves State-of-the-Art Time Series Features

A team of researchers has unveiled OTIS, an open time series encoder designed to provide high-quality features suitable for deployment on resource-limited systems, such as wearables and industrial sensors. Traditionally, the creation of robust general-purpose encoders has depended on scaling laws, which utilize large models to memorize diverse multi-domain training data. This approach, however, poses challenges for practical application due to limitations in memory, energy, and latency. Remarkably, the researchers discovered that adapting standard masked modeling pre-training to the characteristics of time series data results in a compact 7.1M encoder that performs comparably to encoders 54 times its size across 162 tasks, while consuming 10 times less memory, 43 times less energy, and exhibiting 37 times lower latency. The research paper, titled 'OTIS: Learning High-Quality Time Series Features With Tiny Encoders,' can be found on arXiv under the identifier 2410.07299. These results indicate that high performance in time series feature extraction does not always necessitate large scales, potentially enabling deployment on edge devices.