Opening the black box of neural nets: case studies in stop/top discrimination

Abstract

We introduce techniques for exploring the functionality of a neural network and extracting simple, human-readable approximations to its performance. By performing gradient ascent on the input space of the network, we are able to produce large populations of artificial events which strongly excite a given classifier. By studying the populations of these events, we then directly produce what are essentially contour maps of the network's classification function. Combined with a suite of tools for identifying the input dimensions deemed most important by the network, we can utilize these maps to efficiently interpret the dominant criteria by which the network makes its classification. As a test case, we study networks trained to discriminate supersymmetric stop production in the dilepton channel from Standard Model backgrounds. In the case of a heavy stop decaying to a light neutralino, we find individual neurons with large mutual information with $m_{T2}^{\ell \ell }$, a human-designed variable for optimizing the analysis. The network selects events with significant missing $p_T$ oriented azimuthally away from both leptons, efficiently rejecting $t\overline{t}$ background. In the case of a light stop with three-body decays to $Wb\tilde{\chi }$ and little phase space, we find neurons that smoothly interpolate between a similar top-rejection strategy and an ISR-tagging strategy allowing for more missing momentum. We also find that a neural network trained on a stealth stop parameter point learns novel angular correlations.

References in corpus (63)

• Adam: A Method for Stochastic Optimization
• The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations
• An Introduction to PYTHIA 8.2
• End to End Learning for Self-Driving Cars
• DELPHES 3, A modular framework for fast simulation of a generic collider experiment
• Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
• Understanding Neural Networks Through Deep Visualization
• Top++: a program for the calculation of the top-pair cross-section at hadron colliders
• Comparative study of various algorithms for the merging of parton showers and matrix elements in hadronic collisions
• Measuring masses of semi-invisibly decaying particles pair produced at hadron colliders
• Natural SUSY Endures
• Naturalness Constraints in Supersymmetric Theories with Non-Universal Soft Terms
• Visualizing Deep Convolutional Neural Networks Using Natural Pre-Images
• Visualizing Deep Neural Network Decisions: Prediction Difference Analysis
• ZZ production at hadron colliders in NNLO QCD
• Squark and gluino production cross sections in pp collisions at $\sqrt{s}$ = 13, 14, 33 and 100 TeV
• $W^{+}{W}^{-}$ production at hadron colliders in NNLO QCD
• Classification without labels: Learning from mixed samples in high energy physics
• Supersymmetry, Naturalness, and Signatures at the LHC
• QCD radiation in the production of heavy colored particles at the LHC
• t \bar{t} W production and decay at NLO
• Delphes, a framework for fast simulation of a generic collider experiment
• Using Subsystem MT2 for Complete Mass Determinations in Decay Chains with Missing Energy at Hadron Colliders
• Stealth Supersymmetry
• ATLAS Run 1 searches for direct pair production of third-generation squarks at the Large Hadron Collider
• $W^{\pm }Z$ production at the LHC: fiducial cross sections and distributions in NNLO QCD
• t tbar W and t tbar Z Hadroproduction at NLO accuracy in QCD with Parton Shower and Hadronization effects
• Next-to-leading order QCD corrections to t-tbar-Z production at the LHC
• Searching for Directly Decaying Gluinos at the Tevatron
• What is the Machine Learning?
• The Status of GMSB After 1/fb at the LHC
• Top Partners at the LHC: Spin and Mass Measurement
• Decays of the Lightest Top Squark
• Compressed supersymmetry and natural neutralino dark matter from top squark-mediated annihilation to top quarks
• Pileup Mitigation with Machine Learning (PUMML)
• Search for a scalar partner of the top quark in the jets plus missing transverse momentum final state at $\sqrt{s}$=13 TeV with the ATLAS detector
• Measurement of Spin Correlation in Top-Antitop Quark Events and Search for Top Squark Pair Production in pp Collisions at $\sqrt{s}=8$ TeV Using the ATLAS Detector
• Learning to Pivot with Adversarial Networks
• Search for top squarks and dark matter particles in opposite-charge dilepton final states at $\sqrt{s}=$ 13 TeV
• Probing natural SUSY from stop pair production at the LHC
• Search for direct production of supersymmetric partners of the top quark in the all-jets final state in proton-proton collisions at sqrt(s) = 13 TeV
• Closing the stop gap
• Search for direct top squark pair production in final states with two leptons in $\sqrt{s}=13$ TeV $pp$ collisions with the ATLAS detector
• Opening up the compressed region of stop searches at 13 TeV LHC
• Natural Supersymmetry and Implications for Higgs physics
• Probing Compressed Top Squarks at the LHC at 14 TeV
• Very Light Scalar Top Quarks at the LHC
• Soft-gluon corrections for $tW$ production at N${}^3$LO
• Revealing Compressed Stops Using High-Momentum Recoils
• The top squark-mediated annihilation scenario and direct detection of dark matter in compressed supersymmetry
• Four Body Decay of the Stop Squark at the Upgraded Tevatron
• Measurement of the $W^{+}{W}^{-}$ production cross section in $pp$ collisions at a centre-of-mass energy of $\sqrt{s}$ = 13 TeV with the ATLAS experiment
• A top-squark hunter's guide
• Sneaky light stop
• Spot the stop with a b-tag
• A New Method to Visualize Deep Neural Networks
• Playing Tag with ANN: Boosted Top Identification with Pattern Recognition
• Stop Search in the Compressed Region via Semileptonic Decays
• Light stop in the MSSM after LHC Run 1
• Magnifying the ATLAS Stealth Stop Splinter: Impact of Spin Correlations and Finite Widths
• An equal-velocity scenario for hiding dark matter at the LHC
• Compressed Stop Searches with Two Leptons and Two b-jets
• Constraining the Compressed Top Squark and Chargino along the W Corridor

Cited by in corpus (10)

• Learning New Physics from a Machine
• Adversarially-trained autoencoders for robust unsupervised new physics searches
• Machine Learning Uncertainties with Adversarial Neural Networks
• Quark-Gluon Tagging: Machine Learning vs Detector
• A guide for deploying Deep Learning in LHC searches: How to achieve optimality and account for uncertainty
• Mapping Machine-Learned Physics into a Human-Readable Space
• Interpretable Deep Learning for Two-Prong Jet Classification with Jet Spectra
• Per-Object Systematics using Deep-Learned Calibration
• Learning Multivariate New Physics
• Top squark signal significance enhancement by different Machine Learning Algorithms