Reducing sensitivity to systematic uncertainties of the deep neural networks employed in the NOvA experiment

In the NOvA experiment, with its pixelated detectors, measured events can be recorded in an image format. This allows for the usage of powerful image identification techniques such as convolutional neural networks (CNN’s) for the purposes of event classification.

The training data for these networks mostly consists of simulated Monte Carlo data, which closely but not perfectly matches the measured detector data. This leads to the possibility of different performance of the networks on the real data. The differences in the data are thoroughly investigated and quantified in the form of systematic uncertainties.

Here we will utilize the systematic uncertainties to evaluate network performance before deployment, show that including different systematic domains during training can boost both performance and confidence in the networks predictions as well as leverage advances in domain adaption to reduce the effects of systematic uncertainties in the network training itself.