Keywords: generative models, long short-term memory, video, violence
Humans can be overwhelmed by large search dimensionality and are subject to their own biases. This situation can become especially overwhelming in Navy security applications. The purpose of this research is to augment search and detect functions for violent actions in video data. In this talk we will present a deep neural network architecture that has achieved state-of-the-art results on detecting violence in common datasets using 3D convolutional long short-term memory (3DconvLSTM) networks. We will also discuss expanding this architecture to problem of predicting violence. Finally, we will discuss sampling the space of violent video features using deep generative models, for the purpose of assessing model vulnerabilities.