Kumar, Neelesh. Brain-informed deep learning of human movements with neurophysiological interpretations. Retrieved from https://doi.org/doi:10.7282/t3-zhgy-bn04
DescriptionThe accurate and reliable decoding of movement from non-invasive electroencephalography (EEG) is essential for informing several therapeutic interventions, from neurorehabilitation robots to neural prosthetics. However, the EEG's main caveats, namely its low spatial resolution and ill-defined source localization, hinder its reliable decoding, despite progress in both statistical and the most recent machine learning methods. While deep neural networks (DNN) are well-suited for real-time EEG classification, accurate decoding of EEG using DNN requires careful considerations of EEG input representation and network architecture design that can incorporate the spatial and temporal dependencies that exist in the EEG signals. In addition, the network must also be interpretable to highlight correspondence between the learned features and underlying neurophysiology so that their reliability in real-world applications can be guaranteed. In this dissertation, we present brain-informed deep learning solutions for accurate and reliable decoding of movements from electroencephalography (EEG) with neurophysiological interpretations. First, we present a 3-dimensional deep convolutional neural network (3D-CNN) that captures the spatiotemporal dependencies in the EEG in its design and accurately predict complex components of arm movements, namely the reaction time (RT), movement intent, and the direction of the movement. However, the high energy cost of inference using 3D-CNN hinders its use in portable BCI. To address this limitation, we present a neuromorphic solution to EEG decoding that performs as well as the deep neural networks (DNN) while consuming only 5% of the dynamic power, making them suitable for portable BCI systems. Lastly, we develop interpretation techniques to establish correspondence between the features learned by the above networks and the underlying neurophysiology. Our results demonstrate the importance of biological relevance in networks for accurate and reliable decoding of EEG, suggesting that the real-time classification of other complex brain activities may now be within our reach.