They took a few thousand rat neurons and placed them on electrodes. At that point, each neuron's charge could be measured as it fired. A neuron doesn't fire instantly and stop - its charge rises sharply, then falls over time. The graph of a neuron firing resembles a bell curve.
Now that the charge of an individual neuron could be measured, they selected two (one to control the F-22's roll, and one to control the elevation) and began the simulation.
First, imagine a steady, level plane. We say that the plane's error is 0 degrees. Now roll the plane 5 degrees clockwise (or counterclockwise). The error is 5 degrees. So the plane's error is known, and maxes out at 180 degrees (inverted flight).
Second, imagine the graph of a neuron firing: http://img515.imageshack.us/img515/3873/graphet3.png ... We'll say that if we sample from the left side of the graph, we've sampled at X=0 degrees. If we sample from the right side of the graph, we've sampled at X=180 degrees.
Third, a technique exists to deaden a neuron's response. This technique is applied throughout the simulation, so the graph you're looking at is slowly being squished toward the bottom (the dashed line).
Let's say our plane has an error of 10 degrees. Go ahead and sample the graph at X=10. You know what X=10 was at the start of the simulation (the solid line), and you know what it is currently (the dashed line). They use the difference between those values to determine the airplane's corrective action.
Since the simulation starts with the solid line equal to the dashed line, our airplane takes no corrective action and continually plows into the ground. Eventually, the dashed line departs from the solid line enough to cause significant corrective action, and the plane flies straight and level. There is no learning involved, and no decoding of neural communication protocols. They're just decrementing a variable - only the variable is a firing neuron.
It's basically a hi-tech Rube Goldberg device.
You can read the paper here: http://neural.bme.ufl.edu/page13/assets/NeuroFlght2.pdf ... If you can't read a scientific paper, try. It's good to be able to know when a scientist is full of it.
This is no more amazing than putting a rock on a string and telling a computer to correct the plane's roll when the rock moves away from 0 degrees.
Also, the article has the greatest use of hyperbole-by-parentheses ever: "... what scientists are calling a 'live computation device' (a brain)."
palish's comment is informative but not absolute: depending on your interpretation, the headline may not be sensationalist at all. The software was, as described, an "F-22 flight simulator." Confusing to the general public? Yes. Shameful? Not really.
I also disagree that it's a high-tech Rube Goldberg machine. The analogous term for EECS folks would be "hack." It's like saying a wiimote mod for an FPS is a Rube Goldberg machine... which isn't false, but misses the point, don't you think? As a hack, I think it's very interesting. Not surprising, not epochal, but cool nonetheless. A plane can be flown by software, or by bioware, neither approach is perfect at the moment, so "hacks" like these serve their purpose.
But the fact remains, they weren't really using the neurons as a "brain", which is what the article/headline claims, they were just using sophisticated software to find an ewualibrium point and self-correct.
They were hacking the neurons, yes, but only at the most basic level of neuron function. The author misrepresented the facts by suggesting it was "learned" behavior, and that is, I think, shameful.
If the neuron cluster actually learned something, then it would be pretty amazing.