Authors: Kunal Kartik
Abstract: Training convolutional neural networks (CNNs) from scratch using small data sets tend to suffer from over fitting with poor generalization therefore making the models to perform poorly in real world applications. This study examines the effectiveness of transfer learning through the exploitation of pre-trained CNN models and tuning the same to perform classification based on limited data. We measure the performance of popular architectures, including VGG16, ResNet50, and InceptionV3 on several small-scale datasets, including medical imaging and fine-grained object recognition. Systematic layer freezing, targeted fine-tuning, and data augmentation as a part of our methodology are aimed at increasing generalization. As the result shows, training transfer beats training from scratch significantly with fine-tuned models managing to gain up to 25% more accuracy and increased robustness over validation folds. Competitive results were obtained with feature extraction where little fine-tuning was done, which explains its usefulness with limited computational resources. The findings reiterate the value of transfer learning as an applicable solution to small datasets issues, and peers into the best strategies of fine-tuning CNN for data sparse environments.
