Les Houches

This is an old revision of the document!

---

ML: Generative model projects

Discussed the topics we like to study:

• Interpolation and extrapolation properties of generative models and transfer learning

Aims: Define toys cases where the density d(x) is known, Sample training data from this. Then train generative model on training data. Forecast d(x) with that model. Derive pull distribution for the integral of the density in some phase space regions to check how much and if this scales better than sqrt(N) where N are the number of training data points in that phase space regions. Also derive uncertainties via envelope of networks (e.g. via MC dropout, https://arxiv.org/abs/1506.02142) Find out relevant literature.

- Datasets: Toy datasets build on Gaussians, etc. , … madgraph (matrix-element^2, simple process with madgraph) - generative models (Gans, VAEs, …) - try transfer learning (Zee and Zmm) etc.

• Simulator

We have datasample for ttbar,

• Shower development etc.

Information about the projects:

Mailing list: