Turaco: Complexity-Guided Data Sampling for Training Neural Surrogates of Programs

Explore a 19-minute conference talk from OOPSLA2 2023 that introduces Turaco, a novel methodology for sampling datasets to train neural network-based surrogates of programs. Learn how the researchers characterize the proportion of data to sample from different regions of a program's input space based on the complexity of learning a surrogate for each execution path. Discover the program analysis technique used to determine path complexity and see empirical results demonstrating improved accuracy on real-world programs. Gain insights into this innovative approach for addressing the challenge of determining optimal training data for program surrogates, which have applications in various software development tasks. Access the full article, supplementary materials, and related resources to dive deeper into this cutting-edge research presented by experts from MIT and Purdue University.