Brain-computer interface (BCI) is a communication system that translates the brain signal directly to a computer or external devices. It is a promising solution for the patients with neurological disorders as the system is able to restore the movement ability. Various neuroimaging modalities have been utilized for brain signal acquisition, however, functional near-infrared spectroscopy (fNIRS) provides many advantages over other modalities. Hence, it has gained attention for implementing in BCI system. For developing BCI system, the appropriate machine learning algorithm and discriminating features from the hemodynamic response signal are desired, as the previous studies have reported the performance enhancement of fNIRS-based BCI in terms of classification accuracy by focusing on the classifier as well as signal features. The aim of this thesis is to improve the classification accuracy in fNIRS-based BCI by classifying and extracting feature automatically. The convolutional neural network (CNN) was applied owing to the automatic feature extractor and classifier instead of manual feature extraction in the conventional methods. In the experiment, four healthy subjects were measured the hemodynamic response signal evoked by performing tasks including rest, right and left hand motor executions. The conventional methods of fNIRS-based BCI using signal mean, slope, peak, variance, skewness, and kurtosis as the features, and support vector machine (SVM) and artificial neural network (ANN) as the classification methods were compared with CNN-based method. The results show the improvement of classification accuracy of CNN-based method over SVM-based and ANN-based method 6.92% and 3.75%, respectively. The main contributions of this thesis are (1) the promising feature extraction and classification method for fNIRS-based BCI using CNN and (2) the analysis of the feature extracted by conventional methods and convolutional filter of the CNN. ⓒ 2017 DGIST