Context-Aware Intent Prediction for Improved Human-Machine Interactions

Predicting human intentions paves the way of responding to personal requirements accordingly and thus realizes efficient human-machine interactions. Diverse applications, such as human-robot cooperation and smart assisted living, can benefit from precisely predicting human intentions. The electroencephalogram (EEG) signals that reflect electrical dynamics from the human scalp are commonly used to read human's intentions directly. Most conventional approaches to decoding EEG signals heavily depend on engineering experience and domain knowledge and are not able to provide satisfactory performance for widely practical usage. Recently, deep learning technology has shown dominant performance in various fields, and its development for EEG decoding has also sprung up. Considering the limitations of conventional methods for predicting human intentions and the superiority of deep learning technology, in this thesis, I design a set of deep learning frameworks from three aspects of human intention prediction. First, for the subject-dependent application, I design two end-to-end deep learning models with novel spatio-temporal preserving representations of raw EEG data to precisely identify human intentions. I further develop and realize a real-world brain typing prototype with the proposed deep learning frameworks. Besides the end-to-end fashion, I also propose a neural network ensemble model by using a reinforcement learning enhancing scheme. Given that the subject-dependent application requires an adaptation before ready to use on a new subject, I further design a series of models focusing on subject-independent applications. I first develop original deep learning models by leveraging the well-developed self-attention mechanism for effectively capturing the temporal dynamics of EEG streams. Moreover, I design a series of graph representations on top of the attention-based deep learning models to encode nonstructural EEG nodes. Lastly, even though the developed deep learning frameworks demonstrate outstanding performance, they possess the risk of leaking privacy information unintentionally. In view of this gap, I propose a novel user privacy protection framework that transforms raw data into a new format that has a "style" (privacy information) of random noise and a "content" (human intentions) of the raw data. Hence, it is free of user privacy information for training and able to protect all sensitive information at once collectively.