Читать книгу Computational Statistics in Data Science - Группа авторов - Страница 23

Section 3 provides examples of how computational statisticians might address Core Challenges 1–3 (big , big , and big ) for individual models. Such advances in computational methods must be accompanied by easy‐to‐use software to make them accessible to end users. As Gentle et al. [76] put it, “While referees and editors of scholarly journals determine what statistical theory and methods are published, the developers of the major statistical software packages determine what statistical methods are used.” We would like statistical software to be widely applicable yet computationally efficient at the same time. Trade‐offs invariably arise between these two desiderata, but one should nonetheless strive to design algorithms that are general enough to solve an important class of problems and as efficiently as possible in doing so.

Section 4.1 presents Core Challenge 4, achieving “algo‐ware” (a neologism suggesting an equal emphasis on the statistical algorithm and its implementation) that is sufficiently efficient, broad, and user‐friendly to empower everyday statisticians and data scientists. Core Challenge 5 (Section 4.2) explores the mapping of these algorithms to computational hardware for optimal performance. Hardware‐optimized implementations often exploit model‐specific structures, but good, general‐purpose software should also optimize common routines.