EfficientNetB0: Comparing Transfer Learning and Scratch Training on Benchmark Datasets
Rohi Velgina Romould, Prasun Payne, Rajdeep Chatterjee, Mahendra Kumar Gourisaria · 2025
This paper presents a comparative analysis of EfficientNetB0 models trained using transfer learning and from scratch, evaluated across three benchmark datasets: CIFAR-10, MNIST Digit, and MNIST Fashion. The study aims to assess the performance differences in terms of accuracy, precision, recall, and F1-score between the two training approaches. The transfer learning model, pre-trained on ImageNet and fine-tuned for the specific datasets, achieved test accuracies of 80.88%, 98.21%, and 92.71% on CIFAR-10, MNIST Digit, and MNIST Fashion dataset respectively. Conversely, the scratch model demonstrated test accuracies of 87%, 98%, and 91% on the same datasets. Accuracy Precision, recall, and F1-score metrics were further evaluated to highlight the relative strengths and weaknesses of each training method. The results indicate that while transfer learning effectively uses pre-trained knowledge for simpler tasks, training from scratch can yield better outcomes for more complex datasets. This research provides valuable insights into selecting appropriate training strategies for various machine learning applications, contributing to the optimization of model performance in diverse scenarios.