The Self-Simplifying Machine: Exploiting the Structure of Piecewise Linear Neural Networks to Create Interpretable Models

Today, it is more important than ever before for users to have trust in the\nmodels they use. As Machine Learning models fall under increased regulatory\nscrutiny and begin to see more applications in high-stakes situations, it\nbecomes critical to explain our models. Piecewise Linear Neural Networks (PLNN)\nwith the ReLU activation function have quickly become extremely popular models\ndue to many appealing properties; however, they still present many challenges\nin the areas of robustness and interpretation. To this end, we introduce novel\nmethodology toward simplification and increased interpretability of Piecewise\nLinear Neural Networks for classification tasks. Our methods include the use of\na trained, deep network to produce a well-performing, single-hidden-layer\nnetwork without further stochastic training, in addition to an algorithm to\nreduce flat networks to a smaller, more interpretable size with minimal loss in\nperformance. On these methods, we conduct preliminary studies of model\nperformance, as well as a case study on Wells Fargo's Home Lending dataset,\ntogether with visual model interpretation.\n