Astral Codex Ten

fMRI

11 min read

fMRI

Article

fMRI is a recurring concept in the Astral Codex Ten archive, appearing 4 times across 4 issues between April 22, 2021 and July 16, 2024. The archive places it in contexts such as “in a world where fMRIs have only been done on humans, being able to do one on any other animal is a huge win”; “when you put subjects in an fMRI scanner”; “too small to image with standard fMRI methods”. It most often appears alongside Scott, ACX, dopamine.

Metadata

  • Category: Concepts
  • Mention count: 4
  • Issue count: 4
  • First seen: April 22, 2021
  • Last seen: July 16, 2024

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Your Book Review: Are We Smart Enough To Know How Smart Animals Are?
April 22, 2021 · Original source
· Ethology also has some really interesting lessons about how important various practical matters and methodology can be when it comes to what your field can (and can't) produce. For example, it turns out that a surprising amount of useful data about animal cognition comes from experiments with dogs. Is it because dog brains have some interesting physical structures? No, not really that different from a comparably sized mammal. Is it because they are social animals and so have a lot of the same cognitive lego blocks as we do? Maybe a little. The main reason is because they will sit still for an fMRI to be the goodest boy (and to get hot dogs). Turns out sitting still for several minutes in a giant, whirring machine isn't something most animals (including chimps) are that into. So we use dogs.
· Which is super interesting for thinking about experimental design. On the one hand, it shows that you can find some clever "hacks" to find useful data that is otherwise hidden - in a world where fMRIs have only been done on humans, being able to do one on any other animal is a huge win. Lucky for us we've spent the last 30,000 years or so prepping a suitable experimental subject. On the other hand, did somebody say "sampling bias"? How much of what we learn about dog brains will turn out not to generalize like we'd hope? Let's just set aside that dogs have been selected based on ability to empathize with, listen to, and otherwise get along with humans for ~30,000 years, which might result in their cognition being more like ours than other mammals. Even without that, what if we find similarities between humans and dogs that are coincidental and blithely assume they must be common to all mammals? There's a chance we're suffering from the "drunk under the lamppost" effect. Hopefully as we get more and better technology for understanding what's going on in brains (I'm looking at you, Neuralink), we'll find ways to point it at other animals and find out.
Thanks in part to the fMRI studies of dogs mentioned above, we know that not only is a lot of brain anatomy common among mammals, it looks like it's turning on for roughly the same stuff. This means that a lot of what we learn about animal cognition and a lot of what we learn about human cognition can usefully be applied across species (with appropriate experiments to verify, of course). What I find most exciting about this is the fact that it opens up a much wider pool of data for our understanding of not only humans, but also all mammals (and to a lesser degree, all other vertebrates). Findings that might once have been relegated to an obscure animal behavior lab might be tremendously relevant to humans, but nobody bothered to look. As these boundaries get more permeable, we might learn some great stuff.
Highlights From The Comments On Unpredictable Reward
September 30, 2022 · Original source
Most people think this is the case because when you put subjects in an fMRI scanner and have them do tasks where they get reward or learn cues that are associated with reward, you robustly get RPE-related BOLD activity in the NAc, and only rarely/more weakly in the ventral tegmental area (VTA), which contains the dopamine neurons that project to the NAc. So when you see those nice fMRI maps, the NAc is lit up in red. But the physiological basis of this fMRI signal is hotly debated (for example, it could represent primarily synaptic input rather than actual neuronal firing, especially in a GABAergic circuit like the striatum), and in single-unit recordings in mice, rats, and monkeys, it is unequivocal that dopamine neurons in the VTA show much more RPE signaling than the striatum.
Highlights From The Comments On Jhanas
October 31, 2022 · Original source
This 2013 paper, Case Study of Ecstatic Meditation: fMRI and EEG Evidence of Self-Stimulating a Reward System seems to get at the question Scott is posing about which reward machinery is being triggered, using MRI to investigate. In particular they substantiate this hypothesis:
Inline links: Case Study of Ecstatic Meditation: fMRI and EEG Evidence of Self-Stimulating a Reward System
> H5: Jhanas should show increased activation compared to the rest state in the dopamine reward system of the brain (NAc in the ventral striatum and medial OFC). A broad range of external rewards stimulate this system (food, sex, beautiful music, and monetary awards), so extreme joy in jhana may be triggered by the same system (the VTA is also part of this system, but is too small to image with standard fMRI methods, but see [35] for successful imaging methods).
Consciousness As Recursive Reflections
July 16, 2024 · Original source
Brain imaging tech such as functional magnetic resonance imaging (fMRI) does not show thoughts. fMRI shows the slow-changing anatomy of biological neural networks, and how they change on a timeframe of months and years as a nervous system learns, matures and incurs damage. The tiny, milliseconds-long workings of individual neurons and action potentials can also be studied, by isolating them under microscopes and examining the molecules, ions and electrical flows in great detail. Unfortunately, thoughts last from tenths of seconds for the briefest recognitions, to several minutes for the most focused uninterrupted problem-solving. They’re between the timescales that science has great methods for. But a new method can do it! I’ll describe it in the last section of this post, because it provides hope that much of what I’m about to conjecture will be able to be tested experimentally.
Inline links: biological neural networks, neurons
Self-consciousness impedes complex unconscious information processing because it competes with it for neurons. Neurons that are tuned into the rhythm aren’t available for other things, and the recursion of self-referentiality can keep these neurons occupied for a long time. So qualia arise out of neuronal information processing much like biology arises out of chemistry. When chemical reaction chains build each other, they can achieve self-replication. When neuronal activities reflect each other, they can achieve self-reflection. Many processes that know each other become one process that knows itself. From the information processing angle, oscillations that can maintain bits of information have internal working memory, which is the only thing that non-oscillating neuronal activities lack in order to fit the definition of nondeterministic Turing machines. (The IT people among you should grok this immediately. Everyone else may have to dedicate some study time.) From this angle, there are not one, but two levels of information processing systems. The brain is one, obviously. But running inside the brain, oscillations/thoughts with memory are themselves additional information processing systems. It’s analogous to a physical computer system that has, running inside of it, one or more virtual machines. We have failed to locate qualia by imaging the former, because they happen in the latter. This theory of qualia applies only to biological neuronal processes. A for loop is self-referential but is not a biological neuronal process, so I don’t claim it has qualia. “Surely” in the vast space of possible AI architectures, some could be designed to have phenomena that are more or less analogous, but I see no reason to believe the current LLMs do. How to test this theory In the late 1990s and early 2000s, there was much hope in the study of consciousness that then-new functional magnetic resonance imaging (fMRI) tools would let us look into the brain more deeply and thereby let us figure out consciousness. While science did indeed learn much more about the brain, the hope that this would help resolve the puzzle of consciousness did not pan out. But the hope wasn’t crazy: new measuring capabilities are a good reason to expect new data that can hopefully clarify matters. There is new such hope, due to another new method called EEG source analysis. Electroencephalography (EEG) puts electrodes on the scalp and measures tiny electrical currents between them. EEG is very good at temporal resolution, but for most of the century since its invention in 1924, it had almost no spatial resolution. It could tell you the differences between individual milliseconds in your electrical flow measurements, but it couldn’t tell you where in the brain the signals were coming from17. However, if you hook those EEG electrodes up to the amounts of computational power available these days, you can mathematically reconstruct quite good guesses about where in the brain the electrical signals are coming from. And that’s a game changer. This combined temporal-spatial resolution lets you localize individual neural oscillations, if they’re large enough. And that’s how you get to look at (oscillating) thoughts! There are multiple EEG source analysis algorithms. Low-Resolution Electromagnetic Tomography (LORETA) is arguably the best one at the moment. It still has low spatial resolution compared to fMRI, as it says right in the name, and it’ll remain that way. There is a strict physical limit to how much signal this method can ever get out of the noise. But it should suffice for oscillations large enough to exhibit interesting conscious phenomena like self-awareness. Here are a few falsifiable hypotheses that follow from this theory of what makes thoughts/oscillations produce qualia. Most of them could not have been tested without this new method18. Measure EEG of subjects who do the classic test paradigm where stimuli are presented very briefly and subjects report whether they experienced conscious awareness of the stimulus. Run EEG source analysis on the measurements. Variations of the threshold of how long a stimulus has to be presented in order to become conscious should be predictable from variations of the frequency of the oscillation that takes in the signal from the respective sensory neurons.
Inline links: nondeterministic Turing machines, EEG source analysis, 17, Low-Resolution Electromagnetic Tomography (LORETA), arguably the best one, 18
There is another new approach that also improves the intersection of spatial and temporal resolution. It combines the millimeter-scale spatial resolution of simultaneous fMRI and positron emission tomography (PET, that’s the one where you inject a radioactive tracer) with improvements of temporal resolution down to as little as 12 seconds using clever tweaks to radiotracer delivery. Currently that temporal resolution is still too long for most thoughts, but there’s ongoing development and the physical limits to improving the temporal resolution of this method are not yet established. This might end up superior to EEG source analysis, especially for studying the center of the brain.
Inline links: temporal resolution down to as little as 12 seconds