A Resource-Efficient and Data-Restricted Training Method Towards Neurological Symptoms Prediction

Advancement in neuroscience and emergence of deep neural networks (DNN) have significantly improved the performance of neurological symptom prediction. Common practices mainly train models on cloud servers with labeled data and deploy trained models onto device for inferences. However, due to data privacy concerns, troublesome data labeling process and condition variation caused by electrodes shift and physiological status changes, such cloud training and edge inference methodology can hardly be applied to real-life scenarios. In this work, we propose a novel transformer based on-device training framework to tackle the aforementioned issues. The proposed framework adopts direct feedback alignment (DFA) for hardware and computation efficient model training and semi-supervised learning with customized augmented techniques to ease the data labeling process. The effectiveness of the proposed framework is validated on seizure prediction task using electroencephalogram (EEG), experimental results indicated that the proposed framework yields an average improvement of 6.39% and 0.05/h in prediction sensitivity and false positive rate respectively under various label rates.