The goal of this project is to develop deep learning techniques for particle identification and multi-ring event reconstruction in a water Cherenkov detector. It will focus on simulations of NuPRISM and Hyper-K detectors, but may eventually be tested on Super-K data. The project will explore accepted supervised training methods in conjunction with semi-supervised and unsupervised methods such as Variational Autoencoders (VAEs) to enable training directly from data and with limited labeled datasets as well as synthetic data generation. Finally the effect of such limited and biased datasets on systematic uncertainties will be explored and may be mitigated through adversarial techniques. The successful completion of the project will have a substantial impact on key scientific goals of both the Super-K and Hyper-K experiment, such as the CP violation measurement.