CoME: Collaborative Model Ensemble for Fast and Accurate Predictions

Lei Deng, Xinlin Gao, Yi Xiao, Sheng Chang, Xiaochun Cheng, Xianchuan Yu

Abstract: Ensemble learning, which combines multiple weak classifiers to form a strong classifier, has been shown to improve model accuracy and generalization. In the course of our research on convolutional neural networks (CNNs), we have discovered a novel model aggregation method called CoME, which accelerates model training speed, enhances model precision, and exhibits good interpretability. Starting from the relationship between convolutions and planar point classification problems, this paper explores the working mechanism of convolutions and their classification principles. We demonstrate how to leverage these principles to construct our ensemble model. Additionally, during the process of model simplification, we derive the Softmax function and gain new insights into the numerical values flowing through the model, referred to as logarithmic probability. Finally, we conduct experiments to validate the effectiveness of our model.