Review on Transfer Learning for Convolutional Neural Network
Rajdeep Kaur, Rakesh Kumar, Meenu Gupta
Abstract
Convolutional neural network (CNN) has recently received much interest from researchers as an image classification technique. CNN requires a lot of data to train a model from scratch, but in some application areas data is limited which results in an overfitting problem. This study focuses on current research publications that employ Transfer Learning (TL) to overcome the problem of limited data to train the CNN model. TL is a machine learning approach in which the knowledge of the pre-trained model is reused to solve the other problem. Further, this paper presents a review of pre-trained models such as AlexNet, VGG, ResNet, GoogleNet which are used in image classification tasks in different application areas. Two main TL strategies are discussed, first Feature Extractor in which all layers of the pre-trained model are frozen and second Fine-Tuning in which some of the layers are frozen and some are trainable. This review paper focuses on the concepts related to TL and TL strategies used to train the new task models with the help of pre-trained models.