Classifying Neural Signals Related to Action Perception
My Ph.D. dissertation (under advisement from: Andrew Schwartz , Pitt / CMU) was defended on November 11, 2014.
An html version of my defense presentation is available here.
An abridged explanation of one experiment/chapter is explained below:
Neural activity in the motor cortex corresponds to the reach direction of hand. Our lab uses predictive models to decode this cortical signal for controlling external devices such as a robot arm/computer cursor. Signals related to a visualized plan for movement could be useful for prosthetic control.
The cartoon above, depicting a pilot landing a plane, describes two hypothetical movement features that I have found to be represented in premotor cortex. The pilot controls his own arm (red arrow) with a motor command signal. The pilot visualizes the movement of his arm as the movement of the nose of the plane (green arrow). The green movement feature may be an ideal cortical signal to extract, since it corresponds to the visualized goal of the movement.
I developed software (using Java and Java3D) to interface with a stereoscopic "video game" for primates (D3 Visualization). The hand was displayed as a cursor (green sphere), and the goal was to move it to 14 different target locations (blue sphere). My software changed the view of the task so either: a.) the cursor moved in the same direction as the hand (view 1) or, b.) the cursor moved in the opposite direction of the hand (view 2). Our goal was to model any corresponding changes in neural activity. Since the hand was moving the same direction across views, neural activity that modulated the same way across views was most likely corresponding to the direction of the hand. Neural activity that changed across views was most likely corresponding to the movement of the cursor. I designed a statistical classifier based on this premise.
I constructed linear models of neural activity as a function of hand movement direction (using R and Matlab). Resultant coefficients gave each neuron a preferred direction of movement. Neuronal spiking can be considered a stochastic process, so I parameterized bivariate distributions of each preferred direction using a bootstrapping technique. Elliptical boundaries were constructed from the eigenvectors of the covariance matrix. The image above shows two different result-types (neurons A and B) from this analysis. If the prefferred direction changed across views (neuron "A"), I labeled that neuron "cursor-related". Conversely, if the preferred direction stayed the same (neuron "B"), I labeled that neuron "hand-related". I found two significant results from this experiment: a.) neuron types were segregated to different subregions of cortex, and b.) hand-related activity lagged cursor-related activity. This suggests the existence of a serial processing chain from the visualized plan to movement execution. I am currently working on a peer-reviewed publication to report these findings.