Insight Evaluation on Traditional and CNN Features

Truong-Minh Le, Nguyen-Phan-Long Le, Nhat-Gia-Nghi Hua, Nhien-Loc Bui, Vinh Quang Dinh

Abstract: Feature extraction serves as the prerequisite for any intelligent-based applications. There are various methods to extract features from the initial data sets, namely traditional filters and deep learning components. However, current traditional feature extraction methods still have troubles in dealing with automatically updating the weights. This paper will propose a new feature representation method that shows more promising results than any previous methods. This proposed method investigates the benefits of features based on Convolutional Neural Network (CNN) to solve the above problem. CNN’s purpose is to process multi-dimensional data, such as image and time-series data and crucial features are noticed and picked automatically without the need of being supervised by human-being. The proposed method has demonstrated significant improvement in terms of accuracy and stability compared to traditional methods of feature extraction. From that, we can easily understand why CNN-based algorithms have become a part of state-of-the-art algorithms these days.