The paper presents RTSM, a model designed for multivariate time-series classification. Accurate classification in this domain requires extracting temporal features and integrating information across the sequence. The authors propose weight-tied refinement as a mechanism to repeatedly process these features, which offers a way to improve representation without the cost of allocating new parameters.
While the abstract is brief, it highlights the core idea: using a recursive structure with shared weights to refine feature representations over time. This approach could reduce model complexity while maintaining the ability to capture long-range dependencies in multivariate data. The paper is available on arXiv, but full details of the method and experimental results are not included in the abstract.