Meerkat

Meerkat is a technique and software package for multidimensional kernel density estimation.

It uses a variation of kernel density estimation (KDE) technique, relative KDE, where kernel density provides the correction on top of some approximation PDF. The approximation PDF defines the behaviour of the resulting PDF estimate at the edges of phase space and in the region of rapidly-changing structures. This approach allows to use kernel density estimation in a multidimensional case, for the phase spaces with non-trivial boundaries (such as for Dalitz plots).

The technique is described in 2015 JINST 10 P02011 (arXiv:1411.5528).

TensorFlowAnalysis

TensorFlowAnalysis is a tool for performing efficient amplitude analyses using Google TensorFlow library v1.

TensorFlow is an open source library for efficient massively parallel calculations by Google. Its primary use is machine learning, but it is also flexible enough to serve as a general-purpose tool for efficient calculations, in particular, for maximum-likelihood fits with complex models. 

See the slides for the presentation of TensorFlowAnalysis usage. 

Update as of Oct 2020: Since TensorFlow v1 is now retired, and TensorFlow v2 is quite a bit different, the TensorFlowAnalysis package has been redesigned. It is now split in two parts (and has moved to GitHub). AmpliTF is the collection of simple functions and classes useful for HEP calculations. TFA2 is the interface for TensorFlow to perform minimimisation with iminuit, sample random distributions, access ROOT files and simplify plotting data with matplotlib. 

See tutorials how to use TensorFlow with extension libraries for typical flavour physics calculations. 

Model-assisted density estimation with neural networks

ParametricNNDensity is the tool to perform model-assisted density estimation using neural networks. 

The problem of probability density estimation from scattered data arises in many HEP analyses. One example is description of instrumental effects in the kinematic phase space of a multibody decay (such as efficiency profile or background density). In this tool, the neural network is used to learn the functional form of a high-statistics “toy” model of efficiency or a background as a function of a set of effective model parameters, while the sample from detailed simulation or data is used at the second stage to extract the values of the effective parameters. This way, the features of the resulting density can be better controlled by the “toy” model, the ANN learning is more stable since the “toy” sample can be large, and the result is less sensitive to the size of the sample from detailed simulation of the detector since it is only used to constrain a handful of effective model parameters.

The description of the technique can be found in [arXiv:1902.01452]

WIP: Neural nets for efficient toy MC generation