Reviews: Neural Interaction Transparency (NIT): Disentangling Learned Interactions for Improved Interpretability

This paper proposes a novel approach to more interpretable learning in neural networks. In particular, it addresses the common criticism that the computations performed by neural networks are often hard to intuitively interpret, which can be a problem in applications, e.g. in the medical or financial fields. The authors suggest adding a novel regulariser to the weights of first layer of a neural network to discover and preserve non-additive interactions in the data features up to a chosen order and preserve these relationships without entangling them together (e.g. These interactions could then be further processed by the separate columns of the neural network. The approach was evaluated on a number of datasets and seems to perform similarly to the baselines on regression and classification tasks, while being more interpretable and less computationally expensive.