A DNN-Ensemble Method for Error Reduction and Training Data Selection in DNN Based Modeling

Abstract

Deep neural networks (DNNs) have been widely adopted in modeling electromagnetic compatibility (EMC) problems, but the training data acquisition is usually time-consuming through various simulators. This paper presents a powerful approach using an ensemble of DNN s to effectively reduce the training data size in DNN-based modeling problems. A batch of training data with the largest uncertainties is selected using active learning through the variance among the ensemble of DNNs. Subsequently, a greedy sampling algorithm is applied to select a data subset using diversity. Thus, the proposed method can achieve both uncertainty and diversity in data selection. By averaging the outputs of the DNN ensemble, the prediction error can be further reduced. Simple mathematical functions are used to validate the proposed method, and a high-dimensional stripline modeling problem also demonstrates the effectiveness of this DNN-ensemble approach. The proposed method is task agnostic and can be used in other surrogate modeling problems with DNN s.