Abstract
Deep Neural Networks (DNNs) have achieved remarkable accuracy for numerous applications, yet their complexity often renders the explanation of predictions a challenging task. This complexity contrasts with easily interpretable statistical models, which, however, often suffer from lower accuracy. Our work suggests that this underperformance may stem more from inadequate training methods than from the inherent limitations of model structures. We hereby introduce the Synced Explanation-Enhanced Neural Network (SEE-Net), a novel architecture integrating a guiding DNN with a shallow neural network, functionally equivalent to a two-layer mixture of linear models. This shallow network is trained under the guidance of the DNN, effectively bridging the gap between the prediction power of deep learning and the need for explainable models. Experiments on image and tabular data demonstrate that SEE-Net can leverage the advantage of DNNs while providing an interpretable prediction framework. Critically, SEE-Net embodies a new paradigm in machine learning: it achieves high-level explainability with minimal compromise on prediction accuracy by training an almost "white-box" model under the co-supervision of a "black-box" model, which can be tailored for diverse applications.