Statistical learning and inference in neural signal processing

Abstract

Neuromuscular diseases such as brainstem stroke, amyotrophic lateral sclerosis or spinal cord injuries restrict activities of daily living for millions of patients. Such conditions often cause patients severely affected by them to be left in a locked-in state, sustaining loss of voluntary muscle control and restricted communication abilities unless any other means of assistive technology is provided. Brain/neural-computer interface (BNCI) technologies have become one of the most prominent research areas in this regard. Primary motivation of BNCI systems is to provide communication and control means for people with neuromuscular disabilities by establishing a direct brain communication pathway in replacement of peripheral nerves and muscles. Ultimately the capabilities of BNCIs are dependent on the advancements in robust signal processing methods for neural intent inference. Accordingly, neural signal processing is a very active domain of research playing an important role in brain interfacing to facilitate assistive technologies, as well as in fundamental neuroscience to understand the dynamics of the brain. Major challenges in neural signal processing, particularly for non-invasive modalities to monitor brain activity (e.g., electroencephalography (EEG)), are usually caused by the non-stationary nature of the measured neural signals. Our objective in this dissertation is to develop neural signal processing methodologies for non-invasively recorded brain signals that consider beyond heuristic neural feature learning approaches and also account for this stochasticity. We present a collection of work that explores both traditional machine learning based and contemporary deep learning based neural signal processing approaches. Firstly we present a hierarchical graphical model based context-aware hybrid neural interface inference pipeline within an experimental study for multi-modal neurophysiological sensor driven robotic hand prosthetics. Secondly we present an information theoretic learning driven feature transformation concept to extend neural feature dimensionality reduction problems beyond heuristic feature ranking and selection methods. Thirdly we present an adversarial inference approach to learn discriminative invariant neural representations for deep transfer learning in BNCIs, together with neurophysiological interpretability of these invariant deep learning machines. Fourthly we apply this idea in the context of session-invariant EEG-based biometric representation learning. Lastly we present a framework on using generative deep neural network machines to synthesize task-specific artificial EEG signals by manipulating real resting-state EEG recordings.--Author's abstract