Brain-machine interfaces (BMIs) are an emerging field of research that seeks to interface the brain directly with prosthetic devices. Researchers have exhibited compelling proof-of-concept experiments for BMIs, with multiple groups showing impressive demonstrations of non-human primates and humans using neural activity to control external devices. However, significant improvements in both reliability and performance (achieving control and dexterity comparable to natural movements) are still needed for BMIs to be clinically viable.

One of the fundamental challenges in developing BMIs is designing the decoding algorithm (or decoder), which translates the recorded neural activity into a control signal for an external device. Traditionally, decoders were often designed to mimic the natural motor system; decoders would be fitted from neural activity recorded while the subject performed natural movements, so that they could predict or decode the movement parameters from the neural activity. However, such decoders do not take into account an important factor that comes into play when the decoder is used in a real-time BMI system: the feedback that the user receives. BMIs are inherently "closed-loop" systems; users generate neural activity that is transformed into a control signal (via the decoder), and the user receives feedback (e.g. visually), which in turn will affect the neural activity the user generates next. This feedback cycle that occurs during BMI operation can have profound impact on the neurons and the neural activity that is generated.

The first two chapters of this thesis explores differences at the neuronal level between BMI operation and natural reaching. We leverage techniques and insights from network information theory to characterize information encoding in neural ensembles and neural network structure in the two conditions, BMI control versus natural reaching. Using measures of redundancy derived from information theory, we found that neurons directly incorporated in the control of the BMI contained higher and more redundant information regarding the target, compared to surrounding neurons. Using a novel information-theoretic approach for inferring functional connectivity, we found a consolidation of the network connections between the control neurons of the BMI that parallels the learning curve, suggesting the formation of neural circuitry specific to BMI operation. Together, these findings indicate that fundamental changes are occurring at the neural level, which may cause the observed neural activity during BMI control to differ dramatically from during natural reaching.

The subsequent chapter of this thesis describes our progress in building an adaptive BMI system that takes into account the changes in neural activity that occurs in closed-loop, as well as our efforts in improving the long-term reliability of the system. We demonstrate 2-D BMI control using local field potentials as input with closed-loop decoder adaptation (CLDA) to fit the decoder parameters. CLDA is a method of adapting the decoder's parameters that aims to make the decoder's output accurately reflect the subject's intended movements during closed-loop BMI operation. Local field potentials (LFPs) are measurements of the extracellular electric field that reflect contributions from a population of cells. LFPs are an attractive alternative to single- and multi-unit activity in BMIs because they are more robust to signal degradation over time. In addition, our study sheds light on the features of LFP that can potentially be used as input to BMI, with implications for channel and feature selection.