October 17, 2011
Mapping the Mind to Build Smarter Machines
Although the ability of computers to collect and disseminate information has rapidly advanced in recent years, the machines' ability to correctly process and analyze that information—especially from images and photographs—lags far behind. This creates a problem for organizations with too much image data and not enough human beings to sort through it.
A Research and Exploratory Development Department (REDD) team hopes to solve that problem by developing computational strategies, based on human brain structure and function, to quickly and accurately analyze large amounts of information.
Jacob Vogelstein is the lead researcher on the Electron Microscopy (EM) Connectome project, an independent research and development (IRAD) initiative to transform detailed pictures of brain structures into a computational model.
The first step of the project is to create a map of the connectome—the name for the entire web of neural connections in the brain—from high-resolution EM images.
That's not a simple task. With about 100 trillion synapses in the brain, it would be impossible for scientists to manually identify all of a brain's neurons and trace all of their connections. REDD's Jim Burck, systems engineer for the project, estimates that this job would take about 10 trillion hours, or 1.1 billion years, if done by hand. The EM Connectome project will exponentially speed up the mapping process by developing algorithms to trace the neurons, then convert the data into a meaningful representation of how the brain does its computing. In addition to creating novel algorithms, the team collaborates with academic labs at The Johns Hopkins University for some formulas, and uses others available in the public domain.
"For the most part, this project isn't about new math; it's about applying existing math to new problems," says Vogelstein, a program manager for Applied Neuroscience Research and Development.
The team is bringing together experts from across APL: Vogelstein and Burck are working with William Gray, of the Force Projection Department, Angela Hodge, of the Air and Missile Defense Department, and Dean Kleissas and Philippe Burlina, of REDD. The team's ultimate goal is to map the entire connectome, but right now it is focusing on a small but important computing unit, called a cortical column, which is roughly one cubic millimeter in size.
After creating a map of the cortical column, the team will turn its attention to the specific subnetworks of neurons (neural circuits) that process specific functions. "Each neuron basically does the same thing, just like each transistor in a circuit does the same thing," Vogelstein says. "If you want to understand what's happening in the brain, you need to understand the network—the circuit diagram."
Burck explains that the data start as a 3-D image, then get translated into a graph, and finally are analyzed for patterns indicative of circuits. Vogelstein hopes that the process can be used to create models of microcircuits that will mimic brain functions and be used in the next generation of intelligent machines. More specifically, APL's primary goal is to use the connectome information to create better machines for automated image interpretation.
"Any insight into the fundamental modes of computing in the brain would be of great value to all of our sponsors that engage in any kind of computationally demanding program," Vogelstein says. More specifically, he adds, there are a number of government agencies that might be interested in this technology to more quickly triage images collected for intelligence operations.
Vogelstein and the team already know quite a bit about the connectome: They've been working on the Magnetic Resonance (MR) Connectome Automated Pipeline, which examines the brain on a larger scale and looks to identify how different people's brains are wired with regard to cognitive performance and psychological disorders. "The MR project is about creating new tools to assess brain function," Vogelstein says. "The EM project is about creating new insights into how the brain functions."
According to the team, connectomics is an emerging field of science and APL is one of few organizations looking to use it as an enabling technology. Vogelstein describes EM's time horizon as years, not months, and that the initial goal of the IRAD is "partially to spec out how far away we are from being able to exploit this data."
"I am very excited about this project and the associated challenges," Burck says. "It provides opportunities in the areas of massive data storage and retrieval, image processing, machine learning, informatics, and neurally inspired computing. APL brings a fresh eye to each of these, and can also provide systems engineering to put them all together into a cohesive whole."