Measurement Invariance Investigation for Performance of Deep Learning Architectures

Abstract

Models can be compared against each other. Such comparison helps to find strengths and weaknesses of different approaches and gives additional possibilities for model validation. Multi-class classification is a classification task where each image is assigned to one and only one label. Confusion matrix, Precision, Recall, and F1 Score are popular performance metrics. The common sense is that the performance of any architecture is dependent on the sizes of data set. The goodness of the architecture of deep learning models for different data sets is critical issue. This paper implements Pearson correlation coefficient and the multivariate linear regression method to assess the accuracy of deep learning architectures with five performance indicators. The five performance indicators are training loss rate, robustness, training time, number of model parameters and computation complexity. There are five image datasets used to test four deep learning models: Alexnet, GoogLeNet, ResNet, MobileNet to obtain the values of accuracy and other five indicators. The most important contribution of the article is to show that the accuracy indicator related to training loss rate and training time indicators are not dependent on the selection of the data group. According to the definition of Measurement Invariance (MI), the measurement invariance is demonstrated by the linear regression analysis and inner product of the unit normal vector of the linear regression planes.