Portland, Ore. Living neurons taught to control an F-22 flight simulator may yield secrets that could not only demystify silicon-based artificial neural networks but also provide essential clues for understanding and treating neurological conditions, such as epilepsy, that result from malfunctioning neural circuitry.
The mini brain, as its inventor calls it, houses 25,000 rat cortical brain cells on a monolayer atop a 1.6-mm2 chip that interfaces to, nurtures and controls the natural adaptation of the living neural net.
The device is intended to uncover the principles of real neural circuity, for recasting into artificial neural networks as well as for biomedical therapies. It was invented by University of Florida professor Thomas DeMarse, fresh from a post-doctoral stint at the joint Georgia Institute of Technology-Emory University lab of William Ditto and Steven Potter (www.eetimes.com/article/showArticle.jhtml?articleId=18302220), where cultured neurons are routinely interfaced with computers.
"Our ultimate goal is to identify how a neural net computes," said DeMarse, who is now the principal investigator at Florida's Neural Robotics and Neural Computation Laboratory. "We want to extract the rules that it uses to make its biological computations, because the more we know about it, the more we can pour that knowledge into artificial neural networks in silicon."
Thus far, DeMarse said, "we can say that the process is much more like an analog circuit than a digital one: It is able to do everything that it does because it is constantly learning to adapt over time."
By nurturing a monolayer of cortical brain cells atop a grid of electrodes (the cells stay alive from one month to two years and have to be nourished weekly), DeMarse was able to harness a living neural network's natural adaptation to enable it to learn to fly a F-22 flight simulator. The living mini brain was only given control over a 10 percent deviation in the stick position of the virtual F-22, but it was able to learn to fly the plane "true" in about a quarter of an hour.
"Neurons start out independent of each other, but on time-lapse videos you can see them sending out feelers to see what other neurons are nearby. Then, within about 15 minutes, they start self-organizing into an interconnected network that begins to level out and stabilize the plane's flight," said DeMarse, who works with EE Jose Principe, director of the University of Florida's Computational Neuroengineering Laboratory.
"We believe that our applications will not only be artificial neural networks; we are also looking for models of computation that can describe living neural networks well enough to understand maladies like epilepsy," said DeMarse. The latter work has funding from the National Institutes of Health.
At the University of Florida, DeMarse, a trained EE with a doctorate in learning and memory from Purdue, has applied his rich experiences in the lab to experiments with living neural nets. He invented the Neural Robotics and Neural Computation Laboratory's multielectrode array, which enables electrodes measuring just 30 microns, spaced 200 microns apart, to be spread over an area of 1.6 mm2.
The 8 x 8 gold multielectrode array includes traces leading both to and from the electronics that drive it. Both the electrodes and interconnecting traces are etched onto the 1.6-mm2 glass substrate. Over the gold goes a
silicon nitride insulator, with holes etched through it to each electrode.
For the flight simulator experiments, up to 25,000 rat cortical neurons were arrayed in a monolayer atop the array chip. The electronics monitoring the exposed electrodes recorded the activity of the neurons and injected signals into the living neural net.
"This grid is our bidirectional interface: You can listen to what's going on in the population of neurons, and you also have 64 points where you can stimulate just that particular portion of the network," said DeMarse. "The cortical cells are laid down in a smooth two-dimensional monolayer, because that gives us the best chance of being able to record from active neurons. When you start going three-dimensional, your neurons start to get too far away from the electrodes to record from them well."
Since the cortical cells must be fed weekly, DeMarse said, living computers using real neurons would be impractical for widespread use. "No one wants their computer to die if they forget to feed it," he quipped. "This is only a platform for studying neural computation."
The key to harnessing the living neurons to perform useful work in the lab, according to DeMarse, is to insert them into a traditional analog feedback loop: In the lab, the neural network's outputs affect a robotic system, which in turn affects the inputs to the living neural network.
"The grid gives us a real-time feedback loop, so that information from the neurons [controls] a behavior in the robotic system," in this case manipulating the aircraft simulator, said DeMarse. "The neurons control the pitch and roll angle; [they] literally control the stick of the aircraft. When a neural output moves the stick in a particular direction, that changes the attitude of the aircraft. We then take that telemetry information and feed it back into the electrode grid."
To enable the living neural net to learn, the researchers encoded the feedback signal in an analog language that the neurons already understand: frequency modulation. Neural pathways grow fatter if used more frequently, thereby reinforcing the effect of the feedback loop. Conversely, neural pathways that begin to be used less frequently will atrophy and grow thinner, thus attenuating the effect of the feedback loop.
"The trick here is that we can send pulses at any frequency into any of the 64 electrodes in the network, so what we do is let the feedback telemetry modulate the frequency with which we stimulate individual channels in the network's electrodes," said DeMarse. "The feedback stimulation, in turn, produces more or less output, depending on the frequency of the feedback. It's a method of changing the conductivity of the neural circuitry: More stimulation produces more output, and less stimulation produces less output."
In the beginning, the neural network is unconnected and disorganized, so it is as likely to output a large as a small value. This is one reason for giving the neural network control over only a 10 percent deviation from straight and level flight; otherwise, when starting out, it could send the plane into a nosedive or fly it straight up and stall it out. After the neurons begin to self-organize into a living network according to the real-time feedback, however, the plane levels out and flies true.
"Just as in an artificial neural network, we use the feedback to vary the weights until we produce the desired outcome," said DeMarse. "We begin with high-frequency stimulation on all the electrodes, because we know that, over time, [such stimulation] will increase conductivity in the network. Then, as conductivity increases, you get more and more output from the network and that increased response translates into a larger deviation in the control surface, which therefore pushes [the virtual plane] back to stable and level flight."
Once the neural network has learned to fly level by growing its weights, it makes the classic novice's mistake of going for too much of a good thing: It keeps on increasing its weights until it starts veering off course in the other direction. That's when the feedback loop kicks in again. DeMarse's computer translates the data on the errant overcompensation maneuver and decreases the frequency of the feedback in the attendant channel, attenuating its effect until the aircraft is again level.
DeMarse's living neural networks have learned to fly aircraft straight and level even in simulated stormy weather, including hurricane-force winds.
For the future, DeMarse plans to expand from controlling just the pitch and roll over only 10 percent. The group hopes to allow the neural net wider latitude in controlling pitch and roll as well as extend the methodology to control the craft's heading, altitude and other navigational-type functions.
"We are also looking to try the same approach out in a pattern recognition application," said DeMarse. "With 64 channels that we can adjust, we should be able to train [the network] to recognize crude images."
And DeMarse is already working to devise models of computation that can describe living neural networks well enough to expand researchers' understanding of neurological maladies. In that work, he is using complete slices of hippocampus cells from epileptic rats atop his electrode array.
"We think that by studying abnormal systems, like epilepsy, we can learn something not only about how biological computation works but also about how it can go wrong," he said.