Fine-Tuning of Pre-Trained Deep Learning Models with Extreme Learning Machine

Transfer learning allows exploiting what was learned in one situation for faster learning in another situation. It is widely used for object recognition and image classification applications through pre-trained Convolutional Neural Networks models. Different techniques to fine-tune a pre-trained CNN and when to apply each technique are subjects of great research interest. The impact of each technique on the training time as well as test-set accuracy, influences how transfer learning is utilized. In this paper, we evaluate the performance of Extreme Learning Machine verses Fully-Connected layers on overall-training time and testing accuracy, when transfer learning is employed.