-
Brain-machine interfaces (Fagg, Rennaker). Establishing direct connections between
neurons and digital computers provides the possibility for high-bandwidth interaction between
prosthetic limbs and cognitive assistants. One of the key challenges is how to interpret the activity
of many neurons in the context of some task. This project explores a number of representation,
regression, and classification techniques as applied to problems including limb movement (Nemati
et al., 2007; Fagg et al., under review; Fagg et al., 2007) odor encoding and identification (Wilson
et al., 2006; Rennaker et al., 2006, 2007).
- Scientific data mining for computational neuroscience (Lane, Calhoun). Functional magnetic resonance imaging (fMRI) and magnetoencepholography (MEG) provide unprecedented, fine-grained, noninvasive views of the living brain (Kiehl and Liddle, 2001; Clark et al., 2000, 2001; Ranken et al., 2002). Unfortunately, the data from these sensors is also extremely high-dimensional, noisy, and nonlinear (Friston et al., 1995; Buchel and Friston, 1997; George et al., 2000; Clark, 2002; Friston et al., 2003; Penny et al., 2004). In this project, we develop and employ advanced ML methods for analyzing such data. Through Drs. Lane and Calhoun's existing partnerships with the Mental Illness and Neuroscience Discovery (MIND) Institute in Albuquerque, NM, we have access to data from state of the art fMRI and MEG imaging systems and colleagues with extensive neuroscience expertise. A core neuroscience question is to elicit the networks of brain activity underlying a particular behavior or mental illness. Neuroimaging systems report the activity of individual brain regions, but not the interactions among them. There is a substantial hidden-state recovery problem: identify the hidden network behind the observed regional activities (and correlate those networks with behavior or illness (Burge et al., 2004; Qiu and Lane, 2005; Burge and Lane, 2005a,b, 2006, 2007; Clark et al., 2007). In this project, REU students will team with graduate students to do exploratory data mining of neuroimaging data. They will employ tools developed in Dr. Lane’s group and state of the art graph inference tools from the ML community (Chickering et al., 1994; Heckerman, 1995; Getoor et al., 2001; Leslie et al., 2002; McGovern et al., 2003; Friedman and Koller, 2003; Domingos and Richardson, 2004; Jensen et al., 2004; Neville and Jensen, 2004; Wegner et al., 2005; Mahadevan, 2005c,b,a; Maggioni and Mahadevan, 2006) to identify networks in brain data, develop statistical confidence measures on the networks, and will work with neuroscientists to validate the findings.