Reproduce “AutoNumerics-Zero: Automated Discovery of State-of-the-Art Mathematical Functions” 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
Esteban Real, Mirko Rossini, Connal de Souza, Manav Garg, Moritz Firsching, Quoc V. Le, Yao Chen, Akhil Verghese and 2 more
AutoNumerics-Zero, an automated zero-knowledge evolutionary search method, discovers novel and computationally efficient approximations for transcendental functions by optimizing for high-but-finite accuracy. The system produced float32 exponential function approximations that are over 3 times faster than traditional methods while maintaining sub-1 ULP error, with mathematically proven error bounds.
Key insights
- Optimizing directly for high-but-finite accuracy, rather than arbitrary accuracy, allows for the discovery of computationally more efficient approximations tailored to modern digital hardware.
- Zero-knowledge symbolic regression can uncover novel, unconventional mathematical expressions and program structures that effectively reuse intermediate values, outperforming traditional, fixed-form approximations.
- Integrating hardware-specific objectives like ULPs and execution throughput into the search process leads to programs that are not only accurate but also significantly faster on target hardware by exploiting beneficial compilation paths.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.