Many problems, such as in-image alignment, require localized recomputation on sub-inputs to effectively fine-tune a result. This project aims to implement a MapReduce-like parallel framework for computational problems with the following characteristics: (1) A “divide-and-conquer” approach is applicable to solve the problem, whereby the input of the problem is recursively partitioned. (2) Errors in computation cannot be detected until the stage where partitions are being merged back together. In this case, when we find an error while merging, we have to recurse again on the erroneous portion of the output, possibly by running a slower, more accurate version of our computation on the piece, or partitioning in a different way.

“I’ve enjoyed participating in UROPs in the past because it’s a great way to get involved in research and get to work on interesting problems. I’m really interested in algorithms and parallel computation, so I’m excited to get to apply my coursework and experience in those areas to my SuperUROP project.”