Projects
Overview A BCI creates a direct communication channel between the brain and computers, or other external devices, without through the conventional neuromuscular pathway. It recognizes a user's intent by recording Electrophysiological signals, e.g. electroencephalography (EEG) and electrocorticography (ECoG), from the brain with magnetic sensors or other means. The brain signals are then translated into output control commands which reect the user's desire, through a process called decoding.
The existing algorithms, however, tend to be data-driven, ignoring the knowledge specifically related to BCI. By taking advantage of the neurological, physical, and anatomical knowledge, the projects listed below address several fundamental problems: 1). Automatically learn more effective, meaningful and task specific features from raw brain signals; 2). Achieve more accurate and robust decoding by incorporating knowledge about kinematic parameters; 3). Learn more generalizable decoders across both subjects and trials.
Smooth Convolutional Stacked Auto-encoder (SCSA) Most features used by existing decoding models are de ned by expertise. Their optimality for certain BCI tasks is unknown. Furthermore, more objective justification of using these features is still missing. We propose to automatically learn features from raw brain signals with machine learning method. Based on the neuroscience evidence that suggests human brain is organized hierarchically, we propose a multi-layer deep learning structure, which we call smooth convolutional stacked auto-encoders (SCSA), to automatically learn more effective, meaningful and task specific features from the raw ECoG signals.
Decoding of Finger Flexion from Electrocorticographic (ECoG) Signals by
Incorporating Anatomical and Kinematic Constraints The movement of an external device like a prosthetic limb is subject to the related anatomical and kinematic constraints. The existing decoding algorithms ignore these constraints. We propose a Bayesian decoding model to explicitly capture the constraints and to combine them with brain measurements, yielding a significant improvement in decoding accuracy. Specifically, we exploit the constraints that govern finger flexion and incorporate these constraints into a prior model on finger movement. The improved finger exion decoding is then achieved by combining the prior model with the ECoG signals.
Generalizable BCI Decoding with Knowledge Existing decoding algorithms have difficulties to generalize across either trials or subjects. To address this problem, we introduce a weakly supervised learning method, whereby the generic knowledge about the target variables instead of the exact labels are used to train the decoding algorithm. The use of generic knowledge not only improves the algorithm's generalization performance but also eliminates the need of labeling the training samples. The method is then extended to allow online updating of the decoding functions.
Space-Time Varying Dynamic Bayesian Network and Its Application to an
ECoG Based Augmentation System The connections and dependencies among different parts of the brain are important to understand the brain function. However, this important phenomenon is seldom used to assist the decoding. We introduce the spatial temporal varying dynamic Bayesian network (STVDBN) to characterize the spatial and temporal functional dependencies, based on which we construct decoding algorithm.The classifier is successfully applied in an ECoG based augmentation system to predict movement directions.