If you were to look at how a computers' memory is organized, a rough, non-programmer idea of how it would map wouldn't be correct. Big visible things like windows or the start menu aren't put in a single place, etc.
It seems like remarkable that people don't want to imagine there are the rough equivalent of many, many software layers between the level of the neuron and the level of "reason", "emotions", "self", "consciousness" and etc. I think this comes because the "sense of self" is a process that reflexively takes you as primary and indivisible. A framework that view this as generated by lower process violates this.
Or, you know, the analogy of brain as a computer with computerlike memory management simply might not be 100% useful all the time.
> people don't want to imagine there are the rough equivalent of many, many software layers between the level of the neuron and the level of "reason", "emotions", "self", "consciousness" and etc
It's not that people don't want to imagine, it's that you don't need to imagine that. You don't need to force how the brain functions into the model of a modern computer, though it might simplify and be popular. It's often a helpful analogy/metaphor, and for people that know a lot about computers it often allows them to prognosticate and theorize using elaborate analogies built on top of more analogies, but it's equally helpful to not insist that the thing must be the analogy.
For instance, when I talk to a good friend who is a neuroscientist and active in the development of prosthetics with neural interfaces he is far more skeptical (healthily though) of the computational model than any person I speak to in software engineering or the tech industry more broadly. That's likely because the analogy and model can be built upon with further computer analogies, which people who work with computers love. If we were using an automotive model to describe how the brain functions, I'm sure we would have many automotive mechanics theorizing on the nature of cognition too.
It's not that people don't want to imagine, it's that you don't need to imagine that.
I seems like you're hijacking the discussion to your pet issue. My point isn't about how good in general the computer analogy is, it really isn't. You should consider that maps of brain function began quite a while ago, before the start of the 20th (though accelerated by WWI). Here, the analogy was the machine and the mapping of brain followed functional units in machines. And if you consider the point I make (which pretty much echos the article), it's really a counter-example. The multi-layer organization of software show a system doesn't necessarily have to follow naive physical functional units, especially ones we naively perceive. That's it, there's nothing here forcing the computer analogy.
Neuroscientist love to reify a chunk of brain as responsible for function X. They have done this for 160 years. Only Karl Lashley’s work called “localization” into question but his work was swept aside in the Montcastle-Hubel-Wiesel era of big neuroscience.
Now we do toy experiments using optogenetics of a single inbred strain of mouse and delude ourselves into thinking that we are achieving understanding of a highly complex system.
I’ve worked in this field for 40 years and we are not even asking the right questions.
It is a pity neuroscientists do not know more about analog computing. Can a neuroscientist understand an op amp? Probably not.
To share the harsh light—can a CS expert in AI understand how to get to general AI? Probably not unless, like D. Hassabis, you have a solid background in neuroscience.
On one level that’s a reasonable take it, on the other simply having enough data is a prerequisite to come up with the right questions. Astronomers collected literally centuries of data to build up ever more complex epicenter models before ellipsis became an clearly better fit to the data.
IMO, neuroscience simply needs that foundational data and current theory is largely pointless.
Current theory isn't pointless. Current theory straight up doesn't exist.
There is no (non crackpot) theory of mind yet. Most high quality research on the mind (from a non computer science angle) comes from where it breaks down (schizophrenia, autism), since that's where the money is.
Research into AI and neural nets and stuff may change that, but as far as I'm aware an actual model for how thoughts exist doesn't really exist.
The hard problem of consciousness is an impassable obstacle given the set of tools that we have and, arguably, ever can have. If we grant that there is some level of physical description (chemical, atomic, sub-atomic, whichever) at which consciousness best adheres, what do we use to make the connection? Symbolic equations and theorems don't cut it, and that's pretty much all we've got. Physical systems are fully described by the collection of their measurable properties (positions, momenta, charges, etc)--there's no way to connect things going on "outside" with subjective experience.
What would an operator which turns a physical state into conscious experience spit out?
C|state> = ????
Can't write down e.g. "The perception of a red apple on a table". Red? Table? Just symbols. Consciousness and word-symbols are just too far apart.
> The hard problem of consciousness is an impassable obstacle given the set of tools that we have and, arguably, ever can have
I've come to the same conclusion, after reading and thinking about it for 20 years. I no longer search for answers in books, papers, or threads. It's worse than not being able to find any new insights. I never found any insight that goes beyond describing the problem, or defending the existence of the problem. I no longer expect to see any progress on it in my lifetime.
It is more complex than just a one word description but that’s an argument from absurdity.
Let’s suppose we build a teleport device which can make identical copies. As in your standing in pad X then someone else is standing on pad Y that looks identical to pad X and you both respond identically.
Now, whatever that machine reads as your mind state is your actual mind state. We don’t need to make an actual copy to do exactly what you said was impossible.
To clarify, you can’t really capture an image with words but a camera can capture an image as symbols. It’s just that a person looking at those symbols can’t see the image. The same is presumably true of a mind.
I was agreeing with you. And we are arguing from the same side. My comment was more directed at how the meta-opinion in Hackernews generally struggles to step outside of the computational model, or into other possible theories of mind (or memory) developed by philosophers like Wittgenstein, Dennett, or Hacker; and, the result is usually forcing the computer analogy in ways that assumes a kind of blunt physicalism, like the discrete parts of a computer, and often nonsense, as you described. There is often disbelief expressed at the idea that you could use anything but software or computer analogies to describe how the brain functions, or the mind. The assumption is so strong that people do feel forced into the analogy simply because they also understand computers.
There's no reason to think that they compute things the same way silicon "computers" do -- that they're arranged in a von Neumann architecture, or something. But it is true that they perform computation somehow, and therefore are subject to a similar set of constraints, and share similar goals (efficiency, persistence, etc) with modern silicon hardware. (Considering the differences is also insightful; ie our wetware is a much noisier environment than silicon, and likely requires a different approach to error correction.)
This is a useful perspective, for reasons that GP pointed out. We know of many different physical arrangements that are conducive to computing: e.x. Turing machines, von Neumann machines, RNNs are all Turing complete (in principle), and all look very different. So we should question our assumptions about how the brain is organized. Why should we think that, say, sadness "lives" in some physical location in the brain? Does your email "live" somewhere on your computer? (In some sense yes, in some sense, no...).
Is it not equally plausible that the brain implements a very large RNN? And if it does, should we be surprised that if we try to physically locate, say, "sadness", we might be grasping at straws? In the absence of experimental evidence (and even in the presence of it, if flawed assumptions are driving the sorts of experiments we conduct), both seem plausible to me.
Which is just a long winded way to say, I think there is some value in questioning these assumptions. (Not blindly swallowing others, just pushing on why we have the ones we do.)
> I think this comes because the "sense of self" is a process that reflexively takes you as primary and indivisible
Only if you reduce the sense of self to ego function, and a western one at that.
Self is a much more complicated process. And good models of it already acknowledge functional and content heterogeneity eg subpersonalities, embodiment, aspirations, narratives, personhood, boundary problems etc. Self exists because it solves these problems and doing so was adaptive.
> It seems like remarkable that people don't want to imagine there are the rough equivalent of many, many software layers between the level of the neuron and the level of "reason", "emotions", "self", "consciousness" and etc
This is a false dichotomy. We already know we have a grab bag of specialized accelerator “hardware”, but also soft/firm-ware layers that glue things together. That’s the whole point of the article, you can’t encode autobiographical memories without hippocampus but that’s not the only thing hippocampus does nor it is the only thing required to encode those memories.
And people have already upped the ante on this; Cartesian reductionism of trapping “computation” upstairs is also wrong. Cognition requires an embodied and embedded agent. It is not even a mere “brain-thing”.
This reminds me of this nice paper (https://www.biorxiv.org/content/10.1101/055624v2.full.pdf) "Could a Neuroscientist Understand a Microprocessor?" exploring the (nonsensical) results we would get if we'd apply currently used neuroscience approaches to analyze an Atari console running a simple game.
Similarly, there may be a layer below neurons -- microtubules. Something that many neuroscientists and AI researchers don't want to consider seriously, primarily because we're very limited in our ability to examine, interact, and model that layer.
It also throws off by orders of magnitude the effort which would be required to fully model an organism neural behavior.
There may be that layer, but is there evidence for that layer, and I'm not talking about that certain physicists hypothesis about microtubules and free will.
Sir Roger Penrose was investigating non-computability versus computability in brain behavior. I am not sure he even mentions free will.
This highlights some of the unfortunate tensions in this debate, where people may feel obliged to hold to a literal materialism along with a computational theory of mind, and vigorously shoot down anything that threatens to unsettle that somewhat dogmatic position, for fear of strengthening the position of fundamentalist religious believers. In short, it is primarily a political sentiment, not a scientific one.
We should be able to follow where evidence and rational argument leads without being held hostage to such concerns which are strictly irrelevant to the debate.
You are exactly right. I’m a full time geek neurogeneticist. I find most neuroscience models of brain function too neat and simple. My pinned tweet at @robwilliamsiii is one idea that hackers and CS types may enjoy. In brief—where is the clock? Where are the many levels of the stack that you mention.
In memory, sure, the image is in a block (although not really, it's composited (not composted) then shipped over a wire).
But the functionality is basically everywhere. Similar in minds. We can see images "trip" localized circuits when they are recognized, but the comprehension and processing of the scene is muuuuuch more complicated.
Similarly, where is the little 8bit block that handles a click on the start button, and why God Almighty is it so _far_ from the start button image!
It seems like remarkable that people don't want to imagine there are the rough equivalent of many, many software layers between the level of the neuron and the level of "reason", "emotions", "self", "consciousness" and etc. I think this comes because the "sense of self" is a process that reflexively takes you as primary and indivisible. A framework that view this as generated by lower process violates this.