Learning to Generate Fast Signal Processing Implementations
A single signal processing algorithm can be represented by many mathematically equivalent formulas. However, when these formulas are implemented in code and run on real machines, they have very different running times. Unfortunately, it is extremely difficult to model this broad performance range. Further, the space of formulas for real signal transforms is so large that it is impossible to search it exhaustively for fast implementations. We approach this search question as a control learning problem. We present a new method for learning to generate fast formulas, allowing us to intelligently search through only the most promising formulas. Our approach incorporates signal processing knowledge, hardware features, and formula performance data to learn to construct fast formulas. Our method learns from performance data for a few formulas of one size and then can construct formulas that will have the fastest run times possible across many sizes.
Note that this talk is in partial fulfillment of the speaking requirement.