Hacker News Reader

Why would we ever expect qualia to be based on specific and static combinations of neurons?

The belief that particular parts of the brain perform specific functions goes back a long way. One big factor is the commonness of brain injuries during WWI allowed scientists to draw conclusion like "damage in this region results in inability to do this" and that level of understanding has to various extents gone forward since how the brain really works is actually extremely mysterious.

Keep in mind that the scientist could seem localize a function to a group of neurons at a given moment. But you're right, that sort of thinking is too simple.

Luria goes into this question of localization in The Working Brain (a book that's now quite old). Overall, neuroscience has been trying to understand the brain for a long time and not doing very well.

When multiple regions are actually needed to work together to support a function as a network, then this shows up in the lesion and injury data also. There's also a lot of data from surgeries to address severe seizures.

There's some difficulties, such as if the brainstem is also involved in a function, then it's hard to tell because when the brainstem is damaged you're usually too dead to test the function. But I think if you started studying the brain without any prejudices about localization of function, you'd still end up finding a huge amount of it.

Programs are executed in different memory addresses all the time. The hardened linux kernel randomizes itself. It's like a computer alright, just very different from a von-Neuman one and a completely parallelized one. When I started reading the article I expected something mind-blowing, but it always seemed obvious given the sheer number of neurons that they would load balance and feedback, rewire themselves and do all sort of cool, difficult to predict stuff all the time.

I think part of the mystery is that computers can copy binary information losslessly, ad infinitum. So they can move information as often as needed with (almost) no error. We don't think neurons are as reliable, and excessive copying is expected to degrade a memory/representation. It seems like this might work for discrete categorical representations, whose activation is essentially binary, but would gradually fail for anything analog/continuous.

If the computational hypothesis for minds and brains is true, and it is only conjecture though a tempting one, I can only see one way out. Analogies are a wicked thing to indulge in here. But bear with me.

If an engineer designs a computer that computes on encrypted information, from an observer's naive perspective, the entire contents of memory is rewritten with a random value for each step. Yet there is a mathematical relationship between those seemingly random bits and whatever is encoded. Even just using some error correction and non-deterministic parallel processing and it might look pretty much like noise if you don't have the decoder ring.

Something like that is the only explanation I can come up with. Information is stored at a level of abstraction in a structure that is maintained by being continuously re-encoded in the ephemeral computations at a lower level.

If I understood correctly, I think I agree with your loose analogy. I've tried combining some error correction with ongoing re-learning. It stabilizes things against drift, but not indefinitely. This is all in an extremely preliminary cartoon sketch of a model. https://www.biorxiv.org/content/10.1101/2021.03.08.433413v1....

Digital abstractions seem practically worthless when the hardware is not explicitly designed to suppress non binary behavior.

Perhaps analog FPGA's are interesting windows https://news.ycombinator.com/item?id=23432601

The Xilinx 6200 series could safely run randomized bitstreams without damaging itself. Genetic algorithms could be applied to create circuits that worked despite being riddled with what would normally be fatal design flaws.

Considering that mystifying presupposes that the function of the brain is to represent lossless or perfect memories in a static way, rather than say, maintain coherence or consensus across experiences. It's not clear at all that altering or evolving memory is bad for the system overall.

It might very well be a feature of this drift that old memories or information is gradually adapted. After all evolutionary speaking you wouldn't expect the brain to optimize for archiving but for making fitness optimizing decisions in the world.

But human representations do degrade to the extent we forget detail, and also the information changes (sure, the odour might be the same and the experimental controls perfect, but an organism probably should change how they process it over time "new, unfamiliar odour", "smelt that yesterday"... "wonder if anything's happening related to that odour which emerged recently", "meh, background odour") so some of the information update is additive

> I think there is a strong cognitive bias in the present culture of thinking of the brain like a computer or hard drive.

Good point.

The behaviour observed is puzzling if your model says that the mice brains first train on the smell, then once is it learnt the memory becomes "locked in", and then later encounters with the smell are just read-only "recall and recognition" operations. Just like how we might modify a computer's memory and then later use the memory contents in a read-only way.

But even as a layman on the subject, I can see that learning a smell and later trying to identify the same smell isn't two separate kinds of tasks. If I smell a rose for the 1000th time, I'm still learning about its smell. Why would a brain not change?

The whole premise of NeuraLink is built upon this assumption and nobody found it weird...

By the way, I wonder how the NeuraLink team will work around that.

I am guessing that what GP meant was that some people have started thinking of the brain not as a theoretical computer (i.e. as something that can be modeled as a Turing machine or lambda calculus etc), but as a microprocessor. The compational nature of the brain is a pretty wide held belief (though not universal), one that I for one hold.

However, starting to think of the brain as if it should resemble the organization of a microprocessor, with 'ALUs', working memory, well-defined boundaries etc are definitely wrong, and a misapplication of the notion of Turing equivalence. It is nevertheless somewhat common, especially in popculture.

Brain certainly has regions that are structurally different. The logic of differentiation certainly does not correspond to microprocessor, but differentiation is there as a fact.

That interface is based upon learning. It simply means the learning has to be ongoing.

Yes, but that might entail a lot of rework, now that it is a moving target.

Also, that could solve many of their problems. Maybe things did not work out very well initially precisely because of this unsuspectedly huge brain plasticity.

Not sure I understand what problem plasticity poses to NeuraLink. If anything, that is the mechanism that is supposed to power it up. How otherwise brain would learn the meaning of the signals received from it?

NeuraLink was supposed to learn the mapping, yes, but on the assumption that the mapping was relatively stable, unless I am completely mistaken.

What we learn in this article is that the mapping changes a lot, and quickly.

This is an important difference, as previously it could be enough to converge slowly towards the right mapping, but now it has to be done quickly in order to become good enough before starting again a retraining on it.

(To sum up my remark: the retraining has to be fast enough to not become obsolete before it can be used, given the "representation drift".)

You are conflating low-level I/O (like eyes and NeuraLink) with high-level reasoning.

Assume NeuraLink transmits an encoded signal that there's a dog in front of you. Without "representation drift" (or rather its underlying mechanism) I don't think your brain would ever notice that signal has anything to do with dogs. The wire is inserted into a random place, it is not designed to be repositioned to fire up your "dog" neurons it needs to send "dog" signal.

As for the assumption of mapping being stable, there's no such thing at all. First and foremost the brain itself must learn to interpret signals from the wire and send them back. Any mechanism on wire's side would mostly be to establish a feedback loop for the brain to be able to learn to do that (which it will do by rearranging its internal structure).

Wow there is something weird:

I always assumed that the NeuraLink allowed to output from the brain to external interfaces, not to input into the brain.

I have seen the public presentations of NeuraLink by Elon Musk, and the focus was always to extract info from the brain, not the other way around.

Do you have some info related to your interpretation?

Is there a difference between the brain learning input and output? The brain would need to learn that firing that neutron makes the cursor move right.

The presentation with pigs was all about exporting the brain's representation for olfactory stimuli. They were not demonstrating learning to control cursors.

So now I understand why the quiproquo, sorry.

I agree with you that in this situation (cursor controlling) nothing is changed regarding the learning task for the brain.

I think there is a strong cognitive bias in the present culture of thinking of the brain like a computer or hard drive.

I’ve been meaning to coin a fallacy for this. Every era is absolutely certain that fundamental reality is just like [currently dominant technology] and always has been. I expect the fundamental reality of 2200 to be some kind of mystical networked entity, probably either deistic or pantheistic in nature.

Indeed.

I cringe when this mindset worms it’s way into technical terms. I remember cringing when the term ‘genetic algorithm’ became popular. (It’s a great optimization technique for some things, but it has little to do with biological evolution.) Likewise, ‘Neural Nets’. Aaargh... incredible technology, but nothing resembling an actual neuron.

From early decades, the term ‘computer’ — I feel — kinda hampered people’s ideas about what a computer could do. Specifically, the term ‘computer’ or ‘computor’ referred, originally, to a person who added up stuff: it was a job description in the late 1800s and early 1900s. Did the term’s prior usage affect the things for which early computers were used for?

I've been leaning towards a similar mode of thinking, and like to make the point when people get too obsessed about how the universe is a simulation or a neural network.

What I have been missing is a nice list with previous examples of this line of thinking. The only good one that comes to mind is how seafaring and the ubiquity of maps was one such moment in history. I'd be curious if you have other examples.

The Clockwork Universe is my go-to example.

https://en.wikipedia.org/wiki/Clockwork_universe

https://en.m.wikipedia.org/wiki/Mechanical_philosophy

I kind of take the opposite view. I suspect all of the analogies - the clockwork, machine, computer, network universe - may all be right. I think all of these advances and corresponding analogies better help us understand nature, and I think people making these analogies did so for good reasons. The development of the technologies opened the door to new types of system conceptualizations.

I do think these types of analogies don't really work for the brain, though, or at least only scratch the surface. (Clearly the brain networks and computes, but there's much more going on.)

I'm a physicalist, but I suspect it might take longer to fully understand the brain than it'll take to fully understand the universe. (At least to a feasible level of universe detail; the difference is there'll probably be a point where we'll know we know the brain, but we won't necessarily ever be able to know if we've fully cracked the universe at the root, even if we can explain all observed phenomena.)

Sure, I’m not saying that these theories were inherently wrong, just that the idea of the universe being akin to [current technology] is almost certainly incomplete at best and extremely misleading at worst.

> Why would we ever expect qualia to be based on specific and static combinations of neurons?

I agree. In short, they've discovered that they're measuring the wrong thing. Since we don't understand a lot about the brain, that should be unsurprising.

The same thing happens in computers; in many systems, a specific memory cell at a low-level can be easily remapped to different uses over time.

Could it not be a pattern? Either a geometric pattern or a sequence pattern, like Morse code...?

An image is the same image on a different screen; a code stays the same even if produced by different devices.

In the article it seems the word "pattern" is used not in the sense of an actual pattern, ie a shape or design, independent of what produces it, but to indicate a specific group of neurons.

For me (without knowing anything about it), I would think new sensory stimuli are encoded by "smart" neurons, and that gradually, as the stimuli is experienced again and again, less and less "smart" (or versatile, or performant) neurons are tasked with encoding it.

I mean I think it's reasonable to expect that brains function in an evolved way that should minimize energy use when not outweighed by some other benefit. I think it's very valid to wonder what the benefit is for 90% or more of the neurons involved in experiencing a specific, simple stimuli changing over a relatively short time.

yeah, especially in 'getting use' to something. if regular air all of a sudden started smelling like roses, it would be immediately noticeable, but it probably wouldn't be long before we couldn't smell anything different about the air. That has lots of implications on neuron firing, for example, people could have feelings of their significant other, sex, happiness, etc all of a sudden smelling this rosy air.

Indeed, this criticism is as old as Ned Block's Troubles with Functionalism from 1978.

People of note have seriously spoken of "the neuron to detect X".

For stuff like sight things are very rigid. You will have a Jennifer Anderson neuron, that will fire everytime you see her picture.

It is especially weird that people in this day and age would even think of the brain as a single computer instead of as a network of computers. I have no idea how well that analogy holds, but it does seem like a natural base assumption to have. That image you are viewing in your browser might be on a completely different hard drive in a completely different datacenter from month to month.

I don't think it makes sense to think of the brain as even a network of (microprocessor style) computers. The brain is almost surely a computer in the theoretical sense of the word, i.e. it is equivalent to the lambda calculus or a Turing machine.

But there is no reason to expect that the computational model has any resemblance to the one used in PCs. There are good reasons to believe that the brain's functions are not based on binary logic circuits, that (some) memories are not simple storage etc.

Agreed. What I meant was that networking is the paradigm of the times when it comes to computing, so you would expect THAT to be the base assumption, not a 20 year out of date stand-alone computer.

I mean, even modern PCs and phones are networks of computing units internally.