A paper from 2017 entitled ‘Could a Neuroscientist Understand a Microprocessor?’, by Eric Jonas and Konrad Paul Kording is a brilliant example of the kinds of insights that come from interdisciplinary collaborations.
My favourite is the single-lesion study, where the authors studied three different 6502 ‘behaviours’: rendering the Donkey Kong, Pitfall, and Space Invaders video games, while individually removing single transistors and observing if the behaviour would succeed or fail. A biologist friend of mine felt the experimental setup was also typical of the studies often used in genetics, disease, and metabolic investigations.
The results of the experiment were fascinating: while lesioning roughly half the transistors resulted in a total failure to function and the other half seemed to have no impact, a handful of transistors were found to be specific to each game. Certain transistors, when removed, would inhibit the Donkey Kong behaviour, but not impact the Pitfall or Space Invaders behaviour. This could lead a neurobiologist to hypothesise that perhaps there is a ‘Donkey Kong’ transistor: a single device responsible for the ‘memory’ of Donkey Kong.
Other studies from the paper were able to draw more insightful conclusions, such as the clock being two-phase, that the read/write signal is highly significant, that the registers affect the accumulator, or that decoders affect the status bits. However, no study was quite able to thoroughly grasp the higher-level structure and organisation of a microprocessor.
The paper’s thrust was to advance the evaluation and refinement of neurobiology methods and neuroinformatic tools. Perhaps well-characterised devices, like the 6502, could be in silico ‘model organisms’ for refining techniques prior to deployment in expensive in vivo experiments. My intuition tells me that there is much low-hanging fruit to be harvested from further collaborations between hardware reverse engineers and neurobiologists: the premises and techniques used to analyse black-box silicon or undocumented instruction sets might be adaptable to neuroinformatics, thus expanding the toolbox available for decoding the function of our own brains.