Neuroscience Can’t Explain How an Atari Works

When you apply tools used to analyze the human brain to a computer chip that plays Donkey Kong, can they reveal how the hardware works?

Many research schemes, such as the U.S. government’s BRAIN initiative, are seeking to build huge and detailed data sets that describe how cells and neural circuits are assembled. The hope is that using algorithms to analyze the data will help scientists understand how the brain works.

But those kind of data sets don’t yet exist. So Eric Jonas of the University of California, Berkeley, and Konrad Kording from the Rehabilitation Institute of Chicago and Northwestern University wondered if they could use their analytical software to work out how a simpler system worked.

They settled on the iconic MOS 6502 microchip, which was found inside the Apple I, the Commodore 64, and the Atari Video Game System. Unlike the brain, this slab of silicon is built by humans and fully understood, down to the last transistor.

The researchers wanted to see how accurately their software could describe its activity. Their idea: have the chip run different games—including Donkey Kong, Space Invaders, and Pitfall, which have already been mastered by some AIs—and capture the behavior of every single transistor as it did so (creating about 1.5 GB per second of data in the process). Then they would turn their analytical tools loose on the data to see if they could explain how the microchip actually works.

For instance, they used algorithms that could probe the structure of the chip—essentially the electronic equivalent of a connectome of the brain—to establish the function of each area. While the analysis could determine that different transistors played different roles, the researchers write in PLOS Computational Biology, the results “still cannot get anywhere near an understanding of the way the processor really works.”

Elsewhere, Jonas and Kording removed a transistor from the microchip to find out what happened to the game it was running—analogous to so-called lesion studies where behavior is compared before and after the removal of part of the brain. While the removal of some transistors stopped the game from running, the analysis was unable to explain why that was the case.

In these and other analyses, the approaches provided interesting results—but not enough detail to confidently describe how the microchip worked. “While some of the results give interesting hints as to what might be going on,” explains Jonas, “the gulf between what constitutes ‘real understanding’ of the processor and what we can discover with these techniques was surprising.”

It’s worth noting that chips and brains are rather different: synapses work differently from logic gates, for instance, and the brain doesn’t distinguish between software and hardware like a computer. Still, the results do, according to the researchers, highlight some considerations for establishing brain understanding from huge, detailed data sets.

First, simply amassing a handful of high-quality data sets of the brains may not be enough for us to make sense of neural processes. Second, without many detailed data sets to analyze just yet, neuroscientists ought to remain aware that their tools may provide results that don’t fully describe the brain’s function.

As for the question of whether neuroscience can explain how an Atari works? At the moment, not really.

(Read more: “Google's AI Masters Space Invaders (But It Still Stinks at Pac-Man),” “Government Seeks High-Fidelity ‘Brain-Computer’ Interface”)