Electrocorticography (ECoG) records brain activity from the cortical surface. ECoG data analyses has led to significant advancements in neuroscientific research, particularly in two major domains: functional mapping to understand the cortical organization of human brain; and brain-machine interfaces (BMIs) that decode intent from neural data. Designing high performance BMIs is an active area of interest. A discrete BMI design primarily involves decoding specific targets from features extracted from ECoG data. Majority of ECoG based research studies use spectral features i.e. powers in specific frequency bands; which are based on empirical observations. However, given the non-stationarity and variability of neural signals, features extracted in a data driven way could lead to more robust BMIs. In addition to efficient feature extraction and decoder training, the choice of targets presented to BMI user can greatly affect the bit-rate or throughput of the BMI.
The ability to record ECoG data in the order of days in epilepsy monitorig units (EMU); in synchrony with behavioral data through non-invasive sensors like Kinect; has resulted in deluge of large-scale, coarsely labelled datasets. Deep learning architectures are being explored to tackle this big-data problem and extract meaningful spatio-temporal patterns from these data. However, much of the existing research has been relying on architectures that found success in other domains such as computer vision and audio. A more systematic approach towards neural network architecture designs for analyzing large-scale ECoG datasets, that embeds domain knowledge from neuroscience and neurophysiology, is necessary.
The contributions of this thesis are three-fold. Firstly, I show that we can increase the throughput of a speech-based BMI by using targets that generate maximal separation in neural space. Secondly, I show that data-driven features can be learnt in an unsupervised fashion using autoencoders and are more robust compared to linear dimensionality reductions methods like PCA. These features also aid in funcional mapping by identifying functionally similar electrodes in unsupervised fashion. Lastly, I show that cross-subject model can be learnt to decode finger flexions by learning common latent spaces that map activity from different subjects onto a single latent space. By learning temporal dynamics using a recurrent neural network, from this common space, we show that we can decode continuous behvaiors from ECoG.
By analyzing the architectural design decisions on smaller, well-structured and labelled datasets, we can have a smarter approach towards developing deep learning toolkits for larger, coarsely labelled or unlabelled datasets. The methods described in this thesis will aid neuroscientists in ECoG data analysis, clinicians by providing data driven functional mapping methods, and neural engineers by providing more robust machine learning pipelines for BMI design.
