A comparative study on classification by deep learning

The use of feature extraction methods in classification problems improves the performance of machine learning algorithms. In most applications, however, there is no need for deep neural networks to employ such methods. This is because deep neural networks can automatically generate features from raw data in a hierarchical manner. For this reason, this study investigates autoencoder networks which are capable of extracting new features from raw data. Compared with the-state-art-methods, stack autoencoder has demonstrated more efficient performance in classification of the data sets used in this work. Classification results are also justified with statistical analysis which shows that the proposed method yields the best performance among the compared methods.