Transfer learning through deep learning

Deep learning method, convolutional neural network (CNN) outperforms conventional machine learning method in the famous 1000-class large scale image recognition challenge held in 2012. Since then, state-of-art CNN model is getting deeper and its capacity is getting larger. Multiple regularization techniques have been proposed to overcome overfitting, which results in exceeding human’s level. However, every machine learning method is very domain-dependent, an efficient way to increase its ability to generalize well to the real-world domain is still an open research issue. The common scenario where training dataset is small, latency of prediction is critical, or memory resource is constrained would impact the applicability of deep learning. Transfer learning is used to address these restraints. This project focuses on experiment with multiple transfer learning methods through deep learning on a very noisy dataset, namely WebVision Database. In this report, I discuss the current state of deep learning and transfer learning. Then, I review multiple significant transfer learning methods and suggest how we can combine the use of them. I experiment with a baseline transfer method and another adaptation-transfer method on WebVision Database. I find that this dataset is hard-to-transfer and baseline method still convincingly works. I also report on the implementation in detail and evaluate the result obtained. Finally, I explore the potential of further experiment could be done and conclude the project.Bachelor of Engineering (Computer Science