The NFLikelihood: an unsupervised DNNLikelihood from Normalizing Flows

Humberto Reyes-Gonzalez,Riccardo Torre
2024-05-16
Abstract:We propose the NFLikelihood, an unsupervised version, based on Normalizing Flows, of the DNNLikelihood proposed in Ref.[1]. We show, through realistic examples, how Autoregressive Flows, based on affine and rational quadratic spline bijectors, are able to learn complicated high-dimensional Likelihoods arising in High Energy Physics (HEP) analyses. We focus on a toy LHC analysis example already considered in the literature and on two Effective Field Theory fits of flavor and electroweak observables, whose samples have been obtained throught the HEPFit code. We discuss advantages and disadvantages of the unsupervised approach with respect to the supervised one and discuss possible interplays of the two.
High Energy Physics - Phenomenology,Machine Learning,High Energy Physics - Experiment
What problem does this paper attempt to address?
This paper primarily discusses how to use unsupervised learning methods - NFLikelihood based on Normalizing Flows - to solve the learning problem of complex high-dimensional likelihood functions in high-energy physics analysis. The traditional approach is to use supervised learning deep neural networks (DNN) for encoding likelihood functions. The authors demonstrate the performance of Autoregressive Normalizing Flows (ANF) in handling these complex distributions by presenting three different-dimensional physical phenomena likelihood functions (LHC new physics searches, electroweak theory fitting, and flavor physics fitting). They compare the advantages and disadvantages of unsupervised and supervised methods and discuss the possibility of combining the two. The paper first introduces the concept of DNNLikelihood and then provides a detailed description of three physical analysis likelihood functions used for testing, including their parameters and characteristics. Then, the paper proposes evaluation metrics such as the Kolmogorov-Smirnov test (KS-test) and Sliced Wasserstein Distance (SWD) to measure the matching degree between generated samples and real distributions. In the experiment section, the authors demonstrate the application of ANF in toy LHC analysis, electroweak theory, and flavor physics fitting. They adjust the network architecture and hyperparameters to achieve optimal performance. The results show that ANF can accurately learn these complex likelihood functions, including multimodal distributions and complex parameter correlations. Finally, the paper summarizes NFLikelihood as an alternative to DNNLikelihood, emphasizing its advantages in unsupervised learning and pointing out possible future research directions, such as combining DNN and NF likelihood functions and exploring the potential of conditional normalizing flows in learning complete statistical models.