Data-Powered Computing

The traditional view of computing puts the spotlight on the output rather than on the input. The main reason for this -- besides the obvious one, which is that the input is what we have but the output is what we want -- is that computational power is often gained by lowering our expectations about the output side of the computing pipeline (think of randomization, approximation, etc). A shift in focus toward the "input" has occurred lately, however, causing a revolution of sorts in algorithm design. Computing with massive data sets, data streaming, coping with uncertainty, priced computation, property testing, and sublinear algorithms are all parts of the story. I will discuss the past and present of this development and speculate about its future; all in the (biased) context of my own research.