Reducing model bias in a deep learning classifier using domain adversarial neural networks in the MINERvA experiment

Abstract

We present a simulation-based study using deep convolutional neural networks\n(DCNNs) to identify neutrino interaction vertices in the MINERvA passive\ntargets region, and illustrate the application of domain adversarial neural\nnetworks (DANNs) in this context. DANNs are designed to be trained in one\ndomain (simulated data) but tested in a second domain (physics data) and\nutilize unlabeled data from the second domain so that during training only\nfeatures which are unable to discriminate between the domains are promoted.\nMINERvA is a neutrino-nucleus scattering experiment using the NuMI beamline at\nFermilab. $A$-dependent cross sections are an important part of the physics\nprogram, and these measurements require vertex finding in complicated events.\nTo illustrate the impact of the DANN we used a modified set of simulation in\nplace of physics data during the training of the DANN and then used the label\nof the modified simulation during the evaluation of the DANN. We find that deep\nlearning based methods offer significant advantages over our prior track-based\nreconstruction for the task of vertex finding, and that DANNs are able to\nimprove the performance of deep networks by leveraging available unlabeled data\nand by mitigating network performance degradation rooted in biases in the\nphysics models used for training.\n