(A) study on transfer learning of text classification neural network using image data

Abstract

The object of this study is to apply transfer learning of text classification neural network from pre-trained image classification neural network. Although deep learning shows the best performance among machine learning techniques currently, it requires enough amount of training data and bags of time. Moreover, neural network should be constructed every time with different data set, which is cumbersome task. For this reason, this thesis suggests the possible way to train text classification neural network with reduction of learning time and a small quantity of training data by using transfer learning from VGG 16 network which is pre-trained with ImageNet dataset. With regard to input data, text quantization was used with shape of 3-parallel value with certain gap to resembles the structure of image data, rather than with shape of one-hot vector. As a result, not only training time and final accuracy difference between transferred weights and random normalized weights but also, possibility of transfer learning from image to text was confirmed.