Interpreting neural networks is a crucial and challenging task in machine
learning. In this paper, we develop a novel framework for detecting statistical
interactions captured by a feedforward multilayer neural network by directly
interpreting its learned weights. Depending on the desi