Explanations for Neural Networks by Neural Networks

Understanding the function learned by a neural network is crucial in many domains, e.g., to detect model’s adaption concept drift online learning. Existing global surrogate model approaches generate explanations maximizing fidelity between and on sample-basis, which can be very time-consuming. Therefore, these are not applicable scenarios where timely or frequent required. In this paper, we introduce real-time approach for generating symbolic representation of network. Our idea via another (called Interpretation Network, I-Net), maps parameters function. We show that training an I-Net family functions performed up-front subsequent generation explanation only requires querying once, computationally efficient does require data. empirically evaluate our case low-order polynomials as explanations, it achieves competitive results various data complexities. To best knowledge, first attempts learn mapping from networks representations.