[Micand]

As was flagged in the other discussions, see Could a neuroscientist understand a microprocessor: https://journals.plos.org/ploscompbiol/article?id=10.1371/jo...

Or a recap from The Atlantic: https://www.theatlantic.com/science/archive/2016/06/can-neur...

I saw a wonderful recording of a talk the authors gave a few years ago (which I regrettably can't find now), and it was amongst the most eye-opening talks I've ever seen.

[bee_rider]

It is sort of interesting… I think we will all admit that brains are more complex than computer chips. But I still wonder if there’s some general underlying principle that makes brains more reasonable to study in this fashion than computer chips.

For example, they talk about destroying one transistor, observing that the thing can’t play Donkey-Kong anymore, and concluding that that is the DK transistor.

But computer chips are designed to have incredibly long chains of dependencies where each specific transistor does exactly the right thing, every time.

For neurons, it isn’t so specific, right? They all might fire, depending on the timing, and whether or not they are… I don’t know biology, charged or whatever. The whole system works under the assumption that many complements will mis-fire or be duds are any moment.

It seems (to me at least) more reasonable, to come to the conclusion that the DK neuron is really a DK neuron, if removing it causes an un-recoverable DK related failure… because the whole system is based around handling failures! It is somehow special that something can break it.

[jerf]

"For neurons, it isn’t so specific, right? They all might fire, depending on the timing, and whether or not they are… I don’t know biology, charged or whatever."

So we assume. And with some good reason, including both the studies done over the decades, and the fact that we've built systems based on this architecture that match this concept.

However, it is important to observe that this characteristic is still in the metaphorical spotlight. It is a thing that we can discover. If, in fact, there was One Blessed Neuron that contained the most vital aspect of some critical concept in someone's brain, we currently have zero capability to discover it, zero ability to characterize it, and effectively zero ability to manipulate that neuron directly in some experimental fashion once identified. Therefore, we should be fairly suspicious of the claim that we've eliminated this as even being a possibility.

I expect it is unlikely that there is a such thing as the One Blessed Neuron, even so, but there are a large number of other hypothetical organizations that could exist beyond "an amorphous neural net with nothing really located anywhere", and we have good evidence for that as well; the "regions" of the brain, the fact we can visibly see physically different organizations of neurons in certain regions and associate them with certain tasks. I would not even dream of trying to guarantee that there is no structure lying in between the gross differences we already know about and the hypothetical undifferentiated neural mass, the structures in those complexity voids between what we can currently see. We again have a lot of inductive reasons to believe that there is likely more structure there we do not even have a clue exists, on the grounds that every time we get a closer look at something for the first time ever, it is rarely only and exactly what we expected. It's such a notable outcome that it gets called out specifically when it happens, precisely because it's rare.

[rcxdude]

Yes, there's more of a degree of redundancy in brains than microprocessors, but the 'lesion' approach tends to involve affecting more than just one neuron. The main point of that section is that you've got to be careful about what you define as a 'function'. In the processor case, if you defined function in terms of a register or instruction working properly, then this approach could actually give you a pretty good map. But if you've defined it as whether a particular game runs, it doesn't give you a good map at all, because most games exercise most functions of the microprocessor, and so which small parts aren't important to that particular game is only incidental. With brains the main thing is the underlying functions are not really known, so you need to approach with care.

[stochastician]

Author of that paper here -- What an incredibly nice comment to wake up to at 3am, thank you! This paper made me quit neuro and switch to systems that are easier to model (small molecules) which it turns out are still annoyingly difficult!