NumPy and the ecosystem of libraries built on top of it together form one of the most popular numerical computing environments of all time -- but NumPy itself has a number of limitations stemming from decisions made early in its 20+ year development history. I'll review NumPy's internal architecture and outline some ideas for improving it that we plan to implement over the next few years, including better support for alternative storage formats like sparse and out-of-core arrays, richer data formats like categoricals, missing values, and values tagged with physical units, and sketch some ideas for how NumPy could better support just-in-time compilation.
Speaker(s)
Nathaniel Smith
Nathaniel Smith was a computational fellow at BIDS, where he divided his time between computationally informed research on human cognition (esp. language processing) and on building better computational tools for researchers in general.