Sparse Convolutional Neural Networks for NOνA

The NOνA (NuMI Off-axis ν Appearance) experiment measures neutrino oscillations in a nearly pure muon neutrino beam over a 810 km baseline. NOνA has successfully used a Convolutional Neural Network (CNN) as its main event selector and particle identifier for oscillation analysis since 2016. Despite having over 90% efficiency in classifying all neutrino types, our current CNN requires significant GPU resource during training and can only be applied to a window of activity around the beginning of each event. In order to reduce the computational cost, we have implemented a Sparse Convolutional Neural Network (SCNN) which performs convolutions only when the center pixel of the receptive field is non-zero. This reduces the size of training data and improves throughput. First, we have implemented a sparse MobileNet architecture using the Minkowski Engine package and achieved 89% accuracy with 83% decrease in GPU memory used. In addition, we have implemented a sparse FishNet architecture which resulted in training accuracy of ~90%. In the future, we hope to also incorporate semantic and instance segmentation simultaneously into the FishNet architecture for full end-to-end reconstruction.