Abstract
The enduring legacy of Euclidean geometry underpins classical machinelearning, which, for decades, has been primarily developed for data lying inEuclidean space. Yet, modern machine learning increasingly encounters richlystructured data that is inherently nonEuclidean. This data can exhibitintricate geometric, topological and algebraic structure: from the geometry ofthe curvature of space-time, to topologically complex interactions betweenneurons in the brain, to the algebraic transformations describing symmetries ofphysical systems. Extracting knowledge from such non-Euclidean datanecessitates a broader mathematical perspective. Echoing the 19th-centuryrevolutions that gave rise to non-Euclidean geometry, an emerging line ofresearch is redefining modern machine learning with non-Euclidean structures.Its goal: generalizing classical methods to unconventional data types withgeometry, topology, and algebra. In this review, we provide an accessiblegateway to this fast-growing field and propose a graphical taxonomy thatintegrates recent advances into an intuitive unified framework. We subsequentlyextract insights into current challenges and highlight exciting opportunitiesfor future development in this field.