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

04/24/2018
by   Thomas Roxlo, et al.
2

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^ℓℓ, 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 tt background. In the case of a light stop with three-body decays to Wbχ 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.

READ FULL TEXT

page 9

page 11

page 12

page 16

page 17

page 22

page 27

page 30

research
07/09/2018

Spectral Analysis of Jet Substructure with Neural Network: Boosted Higgs Case

Jets from boosted heavy particles have a typical angular scale which can...
research
09/06/2018

Adversarial Reprogramming of Sequence Classification Neural Networks

Adversarial Reprogramming has demonstrated success in utilizing pre-trai...
research
05/03/2023

Bio-Inspired Simple Neural Network for Low-Light Image Restoration: A Minimalist Approach

In this study, we explore the potential of using a straightforward neura...
research
03/18/2014

Similarity networks for classification: a case study in the Horse Colic problem

This paper develops a two-layer neural network in which the neuron model...
research
01/21/2019

Perception-in-the-Loop Adversarial Examples

We present a scalable, black box, perception-in-the-loop technique to fi...
research
04/19/2020

Hedging with Neural Networks

We study neural networks as nonparametric estimation tools for the hedgi...
research
08/13/2023

Separable Gaussian Neural Networks: Structure, Analysis, and Function Approximations

The Gaussian-radial-basis function neural network (GRBFNN) has been a po...

Please sign up or login with your details

Forgot password? Click here to reset