Application of Reinforcement and Deep Learning Techniques in Brain–Machine Interfaces

Abstract

From early adoption in rehabilitation, the brain–machine interfaces (BMIs) have dovetailed into applications empowering humans in controlling external devices such as prosthesis and wheelchairs with a high level of autonomy. The success of such brain–machine interfaces depends on the decoding algorithms that translate the brain activity into the human’s intention or cognition state. Taking advantage of this decoding, a machine can have a robust perception of human’s cognitive state and modify its actions accordingly. This decoding process can be viewed as a machine learning problem where features of brain activities are mapped to some labeled events or classes in a controlled environment. This mapping traditionally relies on the subject and task-specific signal processing approaches. Thus, the conventional machine learning methods fail to generalize well and transfer the learned features between different tasks and subjects, especially in out-of-the-lab applications. Recently, deep learning (DL) has shown great success in learning the patterns from very large data and generalizing well on different applications. With respect to brain activity analysis, deep learning and reinforcement learning (RL) techniques can significantly simplify analysis pipelines and facilitate better generalization between subjects, tasks, and also learn intricate coupled dynamics. In this regard, this paper provides information on the state of the art and challenges in implementing deep learning and reinforcement learning algorithms in brain–machine interfaces. In order to demonstrate the use of deep learning techniques in BMIs, we also present a case study of physical human–robot interaction where the brain activity is used to classify the task difficulty while interacting with the robot.