Reproduce “Programming with Neural Surrogates of Programs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alex Renda, Yi Ding, Michael Carbin
This work introduces neural surrogates, neural networks that approximate the behavior of existing programs to enhance their performance, adaptability, and tunability. The approach, formalized through three design patterns—Surrogate Compilation, Adaptation, and Optimization—is demonstrated on the llvm-mca CPU simulator, achieving a 1.6x speedup, up to 50% error reduction against ground truth, and a 5% accuracy improvement through parameter optimization.
Key insights
- Neural networks can approximate the behavior of existing programs, offering improvements in execution speed, adaptability to real-world data, or parameter optimization.
- The paper identifies three fundamental design patterns for programming with neural surrogates: compilation (speedup), adaptation (matching empirical data), and optimization (parameter tuning).
- A systematic four-step methodology (Specification, Design, Training, Deployment) can guide the development of effective neural surrogate systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.