부산대학교 도서관

Brain-computer interface (BCI) based on electroencephalogram (EEG) using motor imagery (MI) (MIBCI) has been actively studied. This technology is expected to establish rehabilitation for people with disabilities and support wheelchair operations among other benefits. Usually, it is difficult to construct a general-purpose classification model by machine learning because of the large individual differences in MI-BCI. However, the advances in deep learning have increased the classification accuracy of general-purpose learning. Further, when deep learning is used to classify MI, a large amount of data is required for training, which increases the burden on the user. Fine-tuning, which achieves high classification accuracy by relearning and fine-tuning the weights of previously learned models, can be used to reduce the amount of data required for training. In this study, fine-tuning was performed on EEGNet, a deep learning model that can classify MI with high accuracy, using the BCI Competition IV-2a dataset. Further, the required amount of training data and classification accuracy were optimized. Fine-tuning was performed using the pre-training results of eight subjects other than the target subject for training (subject transfer). The target subject displayed a higher-thanaverage classification accuracy as compared to training only the target subject's MI data without fine-tuning (self-training). The average classification accuracy was improved by 5.12% compared to the case where the target subject's MI was used for training without fine-tuning, and the overall average accuracy was about 79.9%. It was further shown that the amount of training data on which the increase in classification accuracy saturates could be reduced by 47.92% on average. Next, the MIs of three healthy male subjects aged 22-23 years (subjects A, B, and C) were measured and fine-tuning was performed on the MI of each of these three subjects to verify the quantity of training data required to improve the classification accuracy. However, for fine-tuning, the weight coefficients of the model obtained because of prior training with MI from nine subjects included in BCI Competition IV-2a were used. Therefore, the classification accuracy was improved by 9.85% to 66.58% on average. In contrast, fine-tuning had little effect on reducing the amount of training data on which the classification accuracy converged. In conclusion, fine-tuning is effective and improves classification accuracy even when the MI is measured under different conditions.