Irregular time series pose a challenge for models that must answer temporal queries at arbitrary coordinates. According to a new arXiv paper, existing approaches typically keep observation-dependent representations that have to be repeatedly resolved for every query, which can be inefficient.
The paper introduces HiPPO Functional Memory, described as offering coordinate-queryable attention for this setting. The abstract frames this as a direct response to the limitations of current methods, but the available text cuts off after the introduction, so the specific mechanism is not detailed in the source.
As a result, the significance here is primarily in the problem statement and the proposed direction: a representation that can be queried directly at arbitrary coordinates, rather than requiring repeated resolution. Further details would need to come from the full paper.