
Hierarchical Ball Attention is a multi-scale local attention strategy designed to capture global+local patterns with linear (or near-linear) cost from large point/time-series/grid-based physical data.
It can be applied to physical simulations, time series analysis, large-scale point cloud vision, and memory-efficient Transformer variants.
arxiv.org
https://arxiv.org/pdf/2502.17019

Seonglae Cho