Methodology for Evaluating the Generalization of ResNet

Abstract

Convolutional neural networks (CNNs) have achieved promising results in many tasks, and evaluating the model’s generalization ability based on the trained model and training data is paramount for practical applications. Although many measures for evaluating the generalization of CNN models have been proposed, the existing works are limited to small-scale or simplified model sets, which would result in poor accuracy and applicability of the derived methods. This study addresses these limitations by leveraging ResNet models as a case study to evaluate the model’s generalization ability. We utilized Intersection over Union (IoU) as a method to quantify the ratio of task-relevant features to assess model generalization. Class activation maps (CAMs) were used as a representation of the distribution of features learned by the model. To systematically investigate the generalization ability, we constructed a diverse model set based on the ResNet architecture. A total of 2000 CNN models were trained on the ImageNet subset by systematically changing commonly used hyperparameters. The results of our experiments revealed a strong correlation between the IoU-based evaluation method and the model’s generalization performance (Pearson correlation coefficient more than 0.8). We also performed extensive experiments to demonstrate the feasibility and robustness of the evaluation methods.