This page is under construction.

Introduction to BEST

BEST is a Neural Network (NN) based AK8 jet tagger, which aims to identify large radius jets originating from boosted heavy objects (H, Z, W, t, b). The BEST method works by boosting Particle Flow (PF) candidates from an AK8 jet into four hypothetical rest frames (H, W, Z, t). In each rest frame, a jet image is produced and high level variables are calculated. The information is used to train the NN.

Super practical information.

The BEST team includes: Robin Erbacher, John Conway, Mike Hildreth, Johan Sebastian Bonilla, Reyer Band, Brendan Regnery, Abhishek Das

Former members: Justin Pilot, Ramya Bashkar

Contacts: Brendan Regnery (brendan.regnery@cernNOSPAMPLEASE.ch), Johan Sebastian Bonilla (johan.sebastian.bonilla@cernNOSPAMPLEASE.ch)

Our project is maintained on GitLab as BEST

• How can one use it.

Useful Links

This includes links to papers, lists of analyses using BEST, and some talks about BEST

Show Useful Links Hide Useful Links

Second Section: Methodology.

How it works

Boosted Event Shape (BES) Variables

* Copy/paste from previous paper.

Boosted Jet Images

To illustrate this idea behind this technique, consider a highly boosted Higgs boson decaying to bb. In the lab frame, this process will create a jet with two separable regions of energy (referred to as subjets). The more boosted the Higgs, the more overlap there will be between these regions. However, in the rest frame these two regions will be back to back.

To accomplish this, we first determine the jet's three momentum and combine this with the invariant mass of a heavy object (H, W, Z, t). Then, this four vector is used to boost the jet constituent particles to the rest frame of that heavy object.

The convolutional neural network identifies patterns in these images, so it is crucial that these images be presented in a way that makes these patterns easily identifiable. This is accomplished with a series of rotations in the rest frame. First, the leading candidate (highest energy jet constituent) is identified and the coordinate system is rotated twice so that the candidate is located at the x-axis. Next, a subleading candidate is identified; this is jet constituent with the highest energy in the region Delta(phi) > 0.3 away from the leading candidate. Then, a third rotation is performed about the new x-axis so that the subleading candidate is in the x,y plane. After completing these rotations, the other jet constituents will be located throughout a sphere in this new coordinate system.

For centuries, cartographers have dealt with problems caused by projecting a sphere onto a rectangle. The lessons they learned provide insight into choosing a useful projection for our images. The boosted jet images use a projection that preserve solid angle and highlight features located along the equator. This is done by creating a two dimensional image with cos(theta) (polar angle) on the y-axis, phi (azimuthal angle) on the x-axis, and color given by the energy contained within that portion of solid angle. Patterns in these images are visible when a large number of jets are averaged over. This is shown below.

click(large), dblclick(back)

Third Section: Performance

Tagging Efficiency (idk...)

• Some plots would be nice.
• A table with benchmarks.

Fourth Section: Technical Info

NN architecture

• Elaborate...
• Why were things chosen.

Input Features

• Maybe a full list of input features.

Fifth Section: Future Improvements and Plans

• Make this twiki
• Point Cloud Stuff

-- JohanSebastianBonilla - 2020-03-28 -- BrendanRegnery - 2020-04-02