ODSL Forum: Transforming the Bootstrap: Transformers for Scattering Amplitudes (Garrett Mertz)

Friday 19 Apr 2024, 15:00 → 16:00 Europe/Berlin

Basement Seminar Room (Origins Building)

Abstract:

Transformers excel on language and image modeling tasks, but still face challenges on numerical tasks. In this talk, I will present an interesting example of an application of Transformers to a challenging numerical problem in theoretical high-energy physics. We use Transformers to predict the integer coefficients of large mathematical expressions that represent scattering amplitudes in planar N = 4 Yang-Mills theory, a quantum field theory closely related to Quantum Chromodynamics. I will first formulate the physics problem in a symbolic-based representation that is amenable to Transformer architectures and training objectives. Next, I will show that a Transformer can achieve high accuracy (> 98%) on two tasks related to coefficient prediction. Finally, I will dive deeper into both the learning dynamics and a number of architecture ablations in order to further illuminate unique aspects of the problem structure.