Neuroscientists use common analytic techniques to decipher three arcade classics—and miss the mark. Here’s what that could mean for brain science.
By Nathan Collins
(Photo: JD Hancock/Flickr)
It might be awfully difficult, but if you set about to reverse engineer a video game, there’s a way to check your work — in theory, you can find the original source code and see how close your reverse-engineered version is to the original. But what if you want to reverse engineer the brain? Neuroscientists have come up with a lot of tools to do that, but no way to check their work — no source code for the brain.
Well, there’s always Donkey Kong, according to a new study—but neuroscientists might not be very happy about what the stubborn ape has to say.
The problem, Eric Jonas and Konrad Kording write in PLoS Computational Biology, is that, although we have increasingly sophisticated methods for understanding the brain, no one’s entirely sure what those methods are telling us. The reason, they argue, is that we don’t have a test brain — one we already understand perfectly — with which to calibrate our tools.
Except, Jonas and Kording point out, we do have a test brain—just not a human test brain. What we have is microprocessors, such as the MOS Technology 6502, which is what’s at the heart of the classic Atari 2600 console. Jonas and Kording’s idea was to try to reverse engineer the 2600 — that is, to try to figure out its circuits and operations, without relying on the original plans — using standard methods neuroscientists use to understand the brain.
Using a computer simulation of the Atari console, the pair conducted a series of studies analogous to standard neuroscience studies. In one set of tests, for example, they first plugged Donkey Kong, Space Invaders, and Pitfall cartridges into a virtual Atari 2600 console. Then, as the games were loading and running, they tracked activity in different areas of its microprocessor to identify functional connections between those regions — exactly the same technique neuroscientists now use to see how different parts of the brain talk to each other.
Their analysis revealed that two decoders — microprocessor regions responsible for uploading information from game cartridges — passed on information to registers and accumulators — the parts of a chip where information is stored temporarily and then processed into output, respectively. That all makes perfect sense, because that’s how exactly microprocessors work: First upload some data, then process it into output for a computer screen or another part of the microprocessor.
But the same analysis revealed some perplexing results. When the Atari 2600 was processing Donkey Kong, the two decoding regions communicated with each other. In Pitfall, accumulators had no communication with registers. The problem: Those conclusions are simply not true. Separate decoding circuits don’t communicate with each other, and if accumulators didn’t communicate with registers, they wouldn’t have any information to process in the first place.
In other words, a standard technique of neuroscience produced results that we know are wrong.
Jonas and Kording are careful to point out that our brains are not microprocessors, and that the methods neuroscientists use today have produced useful findings. Still, there’s reason for caution. “We have found that the standard data analysis techniques produce results that are surprisingly similar to the results found about real brains,” the researchers write. “However, in the case of the processor we know its function and structure and our results stayed well short of what we would call a satisfying understanding.”