FEW-SHOT CLASSIFICATION OF EEG WITH QUASI-INDUCTIVE TRANSFER LEARNING

Abstract: Brain-computer interfaces (BCIs) are devices that enable people with disabilities to use their thoughts to control external devices and restore or improve their bodily functions. One important aspect of BCIs is the classification of electroencephalography (EEG) signals, which measure brain activity and can be difficult to interpret. To address this challenge, we use time-frequency transformations to convert EEG signals into images and employ pre-trained deep convolutional neural networks (CNNs) to classify the images based on whether the subject heard a sound from the left or the right ear. We investigate whether transfer learning, a technique that involves using a pre-trained model on a related task as the starting point for training on a new task, is effective even when the source and target domains are very different. The best classification result achieved was 61.7%, using EfficientNet V2 tuned on 5 different test-subjects, and tuned on the target subject.