Conditional Invertible Neural Networks for HEP

In this talk Johnny will introduce the concept of conditional Invertible Neural Networks (flows), a family of generative models, and present two applications of them in HEP outside of the domain of generative modelling. The first application is v-flows, a regression task trying to recover a degree of freedom which leverages the ability of cINNs to learn a prior distribution from which we can sample. v-Flows infers the momentum of neutrinos produced in collisions without making hard assumptions, instead using our knowledge of the hypothesised process. CURTAINs on the other hand uses the invertibility of cINNs to learn a transportation between datapoints sampled from two different regions of a conditional distribution. This is useful in extrapolation of background data from the sidebands of a distribution (such as the invariant mass) into a signal region and preserving the relationships between the data distribution and the conditional distribution without the need to generate new datapoints by sampling from some base distribution. CURTAINs is applied in the context of a bump hunt for new resonant physics.

 

---

 

https://arxiv.org/abs/2207.00664 https://arxiv.org/abs/2203.09470