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

TL;DR

This paper introduces techniques to interpret neural networks by visualizing their decision boundaries and identifying key input features, demonstrated through case studies on stop particle discrimination in high-energy physics.

Contribution

It presents a novel method for exploring neural network functionality and extracting human-readable approximations, applied to particle physics classification tasks.

Findings

Neurons strongly correlate with known physics variables like $m_{T2}^{ ext{ll}}$.

Networks identify event features such as missing transverse momentum and angular correlations.

Interpretation tools reveal how neural networks differentiate signal from background.

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 $Wb{\widetilde \chi}$ 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.