An Investigation of Cross-dataset Model Generalization of Convolutional Neural Network

Abstract

Transfer learning has become increasingly important as a method to leverage pre-trained models on new tasks, potentially saving significant training time and computational resources. Understanding how well these models generalize to new datasets is critical when people are going to use them to solve real-world problems. This research investigates the transfer learning performance of pre-trained models. Specifically, this work evaluates whether these models trained on dataset Canadian Institute for Advanced Research (CIFAR)-10 can still perform well when transferred to another dataset Self-Taught Learning (STL)-10. Both datasets share the same classes, ensuring a meaningful comparison. The models were trained for five epochs on CIFAR-10 and subsequently evaluated on STL-10. Residual Neural Network (ResNet)18 achieved a maximum accuracy of 41.54% on STL-10, while Visual Geometry Group (VGG)16 reached up to 53.36%. These results show the moderate generalization capabilities of the models and suggest that even though transfer learning is not completely ineffective, there are challenges in achieving high performance on new datasets without further fine-tuning. This study aids in comprehending model generalization and provides insight into the potential and limitations of transfer learning in real-world applications.