Reproduce “Simplifying Polylogarithms with Machine Learning” on OpenResearch

About this paper

Aurélien Dersy, Matthew D. Schwartz, Xiaoyuan Zhang

Researchers at Harvard University developed machine learning models, primarily transformer networks, to automate the simplification of complex polylogarithmic expressions and to reconstruct polylogarithmic functions directly from their symbols. This work addresses a computational bottleneck in quantum field theory calculations, with the transformer approach achieving up to 91% accuracy in symbol integration for polylogarithms up to weight 4, outperforming classical algorithms in some instances.

Key insights

• Machine learning models, especially transformer networks, can effectively learn the 'intuition' required for complex symbolic simplification tasks like polylogarithm identity application.
• Transformer networks demonstrate a strong capability for solving the inverse problem of 'integrating' mathematical symbols back into their corresponding polylogarithmic functions, a task largely unaddressed by classical algorithms for higher transcendental weights.
• A hybrid approach combining classical symbolic methods (e.g., symbol computation) with deep learning for reconstruction offers a powerful paradigm for scientific computation.

Summary by alphaXiv.