轉移學習(Transfer Learning) 是屬於機器學習的一種研究領域，是把已經訓練好的模型參數轉移至新的模型來幫助新模型的訓練，假設大部分數據和任務是存在關聯性的，通過轉移學習可以將模型所學習到的知識透過某種方式來傳遞知識給新模型從而加快並優化模型的學習效率。 EEG 信號的多類分類對於腦機介面 (BCI) 應用至關重要，與兩類案例相比，使用深度學習方法訓練多類模型所需的時間顯著增加。為了加速訓練過程，轉移學習被用來調整從先前訓練的模型中提取的模型參數，以幫助新模型的訓練。然而，由於 EEG 特徵的顯著個體間差異，使用轉移學習對預測模型的泛化受到限制。 本研究共召集4位受試者參與實驗，通過5種運動想像(Motor Imagery)對受試者進行視覺刺激並收集他們的腦電訊號，我們收集了14個電極的時域和頻域訊號，共做了700 次試驗（即每種方向 140 次試驗），並提取他們腦對於五個不同運動方向的特徵，並通過這些特徵進入深度學習分類器進行5類分類。 在這項研究中，我們測試了一種轉移學習方法的信效度，該方法可用於深度學習方法中，將在五個不同運動方向的運動圖像中對多類腦電圖進行快速分類。測量的腦電數據經過標準程序提取時間特徵的ERP和頻率特徵的ERSP，通過轉移學習將模型的泛化能力轉移給其他受試者。通過結果表明，一維頻域和時域特徵在無使用轉移學習的情況Subject-Dependent的準確率分別可以達到88.27%和87.26%，在Subject-Independent的準確率分別可以達到80.67% 和 79.11%。使用二維頻率特徵加上時域特徵在無使用轉移學習的情況Subject-Dependent的準確率最高準確率為86.88%，Subject-Independent的最高準確率則是80.64%。最後使用二維頻率特徵加時域特徵進行轉移學習時，Subject-Dependent的準確率降至73.52%，Subject-Independent的準確率提升至85.86% (訓練時間從110分鐘降至29分鐘)。我們的結果表明Subject-Independent在腦電波多分類上相較於Subject-Dependent會更加困難，並證明對腦電圖多分類使用轉移學習確實能在不降低太多準確率的情況下也減少模型訓練時間。 ;Transfer Learning is a research field of machine learning, which is to transfer the trained model parameters to a new model to help build up the new model. It is assumed that most of the data and tasks are related. The knowledge learned from the pre-trained model can be transferred into the new model in some ways to speed up and optimize the learning efficiency of the model. Multi-class classification of EEG signals is critical for brain-computer interface (BCI) applications, and the time required to train multi-class models using deep learning methods increases significantly compared to the two-class case. Therefore, transfer learning is used to adjust model parameters extracted from previously trained models to aid the training of new models. However, the generalization of predictive models using transfer learning is limited due to the significant inter-subject variability in EEG features. A total of 4 subjects were recruited to participate in this study. The subjects were visually stimulated through five types of motor imagery and their EEG signals were collected. We collected time and frequency domain signals from 14 electrodes. A total of 700 trials (i.e., 140 trials in each direction) were conducted to extract brain features in 5 different motion directions, and these features were input into a deep learning classifier for 5 categories of classification. In this study, we tested the reliability of a transfer learning method that can be used in a deep learning approach to rapidly classify multiple types of EEG images in five different motion directions. The measured EEG data undergo a standard procedure to extract ERP for temporal features and ERSP for frequency features, and the generalization ability of the model is transferred to other subjects through transfer learning. The results show that the accuracy of the 1D frequency and time domain features can reach 88.27% and 87.26% for Subject-Dependent without transfer learning, and 80.67% and 79.11% for Subject-Independent. The highest accuracy of 86.88% for Subject-Dependent and 80.64% for Subject-Independent without transfer learning is achieved using 2D frequency features plus time domain features. Finally, when transfer learning was performed using 2D frequency features plus time-domain features, the accuracy of Subject-Dependent decreased to 73.52% and that of Subject-Independent increased to 85.86% (the training time decreased from 110 minutes to 29 minutes). Our results show that Subject-Independent is more difficult than Subject-Dependent in EEG multi categorization, and demonstrate that using transfer learning for EEG multi categorization can indeed reduce the model training time without much reduction in accuracy.
