Coleman Brings Statistical Modeling Approach to Neuroscience

Todd Coleman's background in computer science and engineering informs his innovative research into neuroscience and development of novel brain-machine interfaces.

As he was finishing his doctoral work at MIT, Todd Coleman's friends from his undergraduate days at the University of Michigan were urging him to put his scientific knowledge and academic credentials to use in ways that would benefit people.

"There were a bunch of people from Michigan who ended up in Boston, at Harvard Medical School or MIT, and almost all of them were working on something that had some sort of biological angle to it," Coleman said. "They were doing M.D./Ph.D.s or doing biological imaging. They all encouraged me to take all this hard core math I was learning for my Ph.D. and apply it toward helping out mankind."

Upon earning his Ph.D. in Electrical Engineering, Coleman was offered a faculty position at the University of Illinois but his Master's and Ph.D. advisor, MIT Professor Muriel Medard, suggested he spend a postdoctoral year in an area different than the disciplines of computer science and electrical engineering he had excelled at during his academic career.

"What I am interested in understanding is how is information about the environment, or information about intent, or information about sound, how is that specifically encoded in the timing of the spike trains and I like to use statistical principles to do that."
- Todd Coleman

So Coleman listened to all the advice and investigated possibilities in the life sciences, discovering a research path rarely taken by electrical and computer engineers: neuroscience. One of the people Coleman talked with was Emery Brown, a doctor and well-known Professor of Computational Neuroscience from MIT's Department of Brain and Cognitive Science and the Neuroscience Statistics Research Laboratory at the famed Massachusetts General Hospital. After meeting with Brown, Coleman decided Medard was right.

"I knew that I was coming here to get the job and (Muriel) strongly encouraged me not to come right away but to postdoc a year in something completely different," Coleman said. "That would give me more time to mature as a thinker, to work on a completely different class of problems. So I decided to try and pursue something biological and that's how I ended up working with Emery Brown."

In Brown, Coleman found a researcher who is considered a leader in the area of mathematical modeling of neural systems and a mentor he could look up to.

"I had already chatted with him once so I knew him and heard very good things about him, heard people called him on a first name basis," Coleman said. "He had a very down to earth personality and was a first rate scientist and worked on very interesting, cool problems at the intersection of statistics and biology. So I noticed there was a natural opportunity for mentorship and as I got to know him it just naturally evolved in that direction. I would love to be on his level someday; that's quite something to aspire to."

Coleman is well on his way toward that goal. As a member of the Beckman Institute's Artificial Intelligence group and Assistant Professor of Electrical and Computer Engineering at Illinois, Coleman's research focuses on more traditional computer engineering topics but also includes, thanks to his postdoctoral stint at MIT, a large measure of computational neuroscience. He is applying his computer science and electrical engineering skills to neuroscience by using statistical and computational approaches toward understanding brain function and toward the development of applications like novel non-invasive brain-machine interfaces.

The work involving neuroscience could end up satisfying his friends at MIT who were urging him to take on work that would benefit people. Applications from Coleman's research in computational neuroscience could lead to applications such as a brain-machine interface that guides a prosthetic limb.

Coleman said it was during his post-doctoral work with Brown that he studied the probabilistic structure of how neurons in the brain communicate based on what are called spikes and discovered that the same theory could be applied to artificial systems.

"Any neuroscientist is trying to understand how the brain represents information," Coleman said. "There are lots of ways that one can think about doing that and the rubric that I am using primarily, which I learned in my postdoc, is by virtue of how the neuron spike trains encode information."

Coleman said that process "involves statistical reasoning with point processes.

"We know that neurons generate these little flickers of energy called action potentials and it's basically the timing at which they generate all these spikes that is carrying all the information," he added. "What I am interested in understanding is how is information about the environment, or information about intent, or information about sound, how is that specifically encoded in the timing of the spike trains and I like to use statistical principles to do that."

Coleman's research focus of using rigorous statistical approaches to categorize brain function is fairly unique.

Along with his students, and collaborators and Beckman colleagues Tim Bretl and Ed Maclin, Coleman also addresses brain-machine-interface problems by using non-invasive EEGs to record the brain's electrical signals of test subjects during task performance. Coleman's methods use adaptive querying techniques, inspired by data compression, optimal control, and feedback information theory principles that explicitly take into account the user's behavior. This approach - as opposed to most current brain-machine interfaces - includes the user's brain as part of a dynamic system. Coleman said this perspective was strongly influenced by his collaboration with Bretl, an Assistant Professor of Aerospace Engineering who is an expert on dynamical systems.

"The first component is just to understand brain function by virtue of these neural spike trains and the second component of my interest is - remember I'm an engineer originally - that what I would like to do ideally is to understand the brain, but then use that to engineer some sort of system," Coleman said. "What I'm interested in are engineered systems where the brain is in the loop.

"So we use the EEG equipment to record these brain signals and then 'decode' them, to operate some sort of device, be it a cursor on a screen or some sort of actual device with dynamics. Then the user, by virtue of the visual feedback from what he sees, is going to control this process by what he thinks. We take these EEG signals and we perform statistical modeling and signal processing algorithms to develop beliefs, or likelihoods, about what the 'intent' was. Then once we have these likelihoods, we use them to drive some sort of signal, be it a cursor or text or something else. Lastly, we provide feedback to the user - by virtue of the actuation, and also by extra visual information - about our belief in his or her intent. This closes the loop. Tim and I spent a long time thinking about how to model this, and we hope we're onto something."

Coleman's approach is novel in that it explicitly models the interface in a closed-loop, dynamical fashion.

"People have used EEG to control cursors, to control text, to spell a sentence, or type an email," he said. "There are others who have actually used neurons from the motor cortex, they have used invasive technologies, to record those neurons for more sophisticated things like robotic limbs. But most of the approaches are implemented from a more feed-forward perspective."

Coleman said designing and fashioning a closed-loop dynamical system with feedback is a challenging task for several reasons.

"First of all, trying to characterize what the 'optimal' thing to do in lots of scenarios is just way too complex from any theoretical standpoint," he said. "Second of all, even if we could do that, it would require a good model of human behavior and human cognition - a cognitive model of how a user reacts, based upon what the users intent is, to feedback that he sees. That requires some sophisticated cognitive modeling.

"So it's an interesting interplay between principles of control and engineering in what's called information theory, conveying the users' intent, as well as the cognitive modeling of how the user reacts to what he sees. It's a non-trivial task for a number of reasons but we're chipping away at it slowly but surely."

Coleman said developing advanced prosthetics is just one potential application.

"We are trying to think more broadly than that," he said. "Chatting with Tim and Ed has been great in that regard as well - they both really think outside the box. Our collective, complementary strengths provide vibrant discussions during our joint group meetings - discussions and directions to pursue that I don't think any one of us could have by looking at these problems individually. Just the whole idea of having the brain in the loop for these brain-machine interfaces, in a broad sense, opens up so many different opportunities. Prosthetics is one angle. But there are many, many, more."

Coleman is a native of Dallas who went to Michigan on a scholarship. He wasn't sure if he wanted to pursue advanced degrees in computer science or electrical engineering when he went to MIT, but he said having Medard as an advisor led him to work on interesting information theory problems. His relationships with Medard, a former Illinois professor, and Brown remain two of the most important to him.

"Emery and Muriel are two of my closest mentors now," Coleman said. "Both of them are down-to-earth, normal people who value family and treat people as human beings. Everyone calls them on a first name basis - the human element is very strong in both of them - and I admire them for that."