Scaling laws for amplitude surrogates

Value of the Information & Strength of the Argument

The talk provides valuable insights into the empirical scaling behavior of neural network surrogates for scattering amplitudes. The argumentation is solid, based on systematic experiments across many processes and resources. The speaker clearly explains the methodology and acknowledges limitations, such as the need for learning rate optimization. The discussion of the relationship between scaling exponents and degrees of freedom is interesting, though not fully conclusive. The presentation is well-structured and the results are convincing.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high: the speaker uses established tools (MadGraph) and clearly defines the setup. However, no external sources are cited in the talk, and the description provides no links. The title accurately reflects the content. The speaker mentions a paper by colleagues but does not provide a reference. The lack of citations limits the ability to verify claims, but the methodology appears sound.

The title accurately reflects the content, which focuses on scaling laws for machine learning surrogates of scattering amplitudes.

8/10

The talk presents original research with clear methodology, systematic scaling analysis, and comparisons across many processes. The speaker is transparent about limitations and ongoing work. However, the lack of published paper or external references limits verification.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and motivation: scattering amplitudes are expensive, ML surrogates show promise.
• 02:00Scaling laws in deep learning: power-law improvement with data, compute, parameters.
• 05:00Goal: determine universal scaling laws for amplitude surrogates across many processes.
• 08:00Data generation with MadGraph, wide range of processes.
• 10:00Setup: neural network architecture, training details, resources.
• 12:00Results for QCD to top pair: clear scaling laws and plateaus.
• 15:00Scaling exponents and relation to degrees of freedom.
• 18:00Heteroscedastic loss for uncertainties, calibration and scaling.
• 21:00Results for other processes, universality of scaling.
• 24:00GATr architecture for permutation invariance, promising results.
• 27:00Conclusions and outlook.
• 30:00Q&A: discussion on early stopping and learning rate schedules.

• Scaling Laws for Neural Language Models — Provides the theoretical and empirical basis for scaling laws in deep learning, which the talk extends to amplitude surrogates.
• Neural Scaling Laws — A broad study of scaling laws across domains, supporting the universality observed in the talk.