No, Brains Have Nothing Like a Large Language Model (LLM)

Quanta Magazine is a widely-read online magazine with slick graphics. On topics of science the magazine again and again is guilty of the most glaring failures. Quanta Magazine often assigns its online articles about great biology mysteries (involving riddles a thousand miles over the heads of PhDs) to writers who lack even a bachelor's degree in biology, and who may also lack any history of writing very much about biology. Often it will assign such articles to be written by people identified as "writing interns."  The articles at Quanta Magazine often contain misleading prose, groundless boasts or glaring falsehoods. I discuss some examples of such poor journalism in my posts here and here and here and here.

The writers at Quanta Magazine often sound like the most credulous pushovers for scientists making dubious boasts. They often seem like the science journalists depicted below:

A recent  example of a puff-piece article in Quanta Magazine is its fawning article entitled "The Polyglot Neuroscientist Resolving How the Brain Parses Language." The article title is a very misleading one. It is not brains that parse language. It is people who parse language. Parsing language (interpreting what someone meant upon hearing or reading something) is an example of a very high-level mental faculty.  Neuroscientists have no credible tale to tell of how a brain could produce such a faculty. 

The puff piece article is written by someone identified as a "writer and filmmaker," who has written many articles at the Quanta Magazine dealing with topics related to AI and technology, but almost no articles at the site relating to psychology or neuroscience.  What we get is an article devoted to glorifying neuroscientist Ev Fedorenko. You can tell what a "going gaga" scientist glorification affair is occurring by the fact that the article has five huge photos of Fedorenko, each of which fills up my computer monitor. 

We have this extremely misleading statement by the neuroscientist, not matching anything ever found in a brain:

" 'You can think of the language network as a set of pointers,'  Fedorenko said. 'It’s like a map, and it tells you where in the brain you can find different kinds of meaning. It’s basically a glorified parser that helps us put the pieces together — and then all the thinking and interesting stuff happens outside of [its] boundaries.' ”

The brain has no maps, and it has no pointers. I know pointers well, having used them extensively when I was a C programmer early in my programming career. A pointer is a variable that stores the address where some data is stored. Brains have no addresses, and nothing corresponding to a programming variable. So in a brain there can be nothing like a pointer. 

The article has a reference to a scientific paper by Fedorenko, one entitled "The language network as a natural kind within the broader landscape of the human brain." The paper says, "In this Review, we discuss brain areas that are specific to language —what we refer to as the language network — and position them in relation to perception, motor planning and cognition (Fig. 1)." There are no regions in the brain "specific to language" in the sense of being able to parse language. 

The paper announces that it will back up such claims largely by appealing to brain scan studies, saying, "We primarily draw on fMRI data from studies that have relied on the individual-subject functional localization approach14,19 (Box 1), which was essential in clarifying the distinctions discussed.." We then have in Figure 1 an extremely dubious visual showing five brains. We have some "function localization" claims:

• One brain visual has a small fraction of it colored blue, and this is labeled as the "perception" area of the brain. 
• Another brain visual has a very small fraction of it colored red, and we are told that is the "motor planning" area of the brain. 
• Another brain visual has five parts of it colored purple, and we are told that this is the "language" areas of the brain. 
• Another brain visual has quite a few parts of it colored green, and we are told that this is the "knowledge and reasoning" areas of the brain. 
• Another brain visual has quite a few parts of it colored green, and we are told that this is the "intended meaning" areas of the brain. 

People have been making these kind of functional localization claims about the brain for almost 200 years, and almost always the claims have been very dubious. For example, below is an illustration from the beginning of the 1834 book A System of Phrenology by George Combe. Notice how the brain areas have little numbers next to them, with the bottom legend explaining the claims made about localization of function in the brain. 

What basis does Fedorenko have for the brain function area maps in her paper, Figure 2, one rather reminding you of the phrenology image above? She says her basis is brain scan data.  The Supplementary Data part of the paper promises that if you click on a link you will get the data that was used to generate the brain function area maps. Following the links takes you to a little page that is almost worthless for inspecting data, as it involves a .zip file consisting of files in an .nii format that the average person will be unable to load or read. Consequently it is all but impossible for anyone to check the evidence basis for Fedorenko's brain function area maps, which should not be regarded as reliable. Neuroscientists dramatically disagree with each other when they produce such brain function area maps, which typically fail to have a sound evidence basis. 

A figure on page 294 of the paper  gives us the bar chart below: 

The purple bars at left are probably incorrect

The scale on the left shows us that the bars refer to fMRI percent signal changes in different regions of the brain during different activities. You should note well that the largest bars show a variation of only 2%. For almost all of the activities listed, the percent signal change is less than 1%. On average the change in brain signal strength is merely about 1 part in 200. A change that small is no real sign of brains working harder during some cognitive activity. We might expect that mere chance variations would produce differences that small. 

Except for the two purple bars on the left-most edge, the diagram is consistent with what I have often stated on this blog. For example, in my post "Brain Imaging Shows No Appreciable Neural Correlates of Memory Activity," I quoted quite a few studies showing a variation of only  1 part in 200 when people had their brains scanned by fMRI scanners while they were doing various memory-related activities. I noted that so small a change fails to provide any good evidence of brains causing mental activities, as we might expect a 1 in 200 fluctuation in brain activity to occur by chance, even if brains don't make minds. 

But what about the purple bars on the left of the chart? What is the basis for the claim being made in the graph that during sentence comprehension there is up to a 2% change in brain signal strength? We fail to get a justification for the data shown. The references in the paper include some papers referring to sentence comprehension. But none of them seem to back up the claim above. Specifically:

• We have a reference to a paper "Cognitive control and parsing: Reexamining the role of Broca’s area in sentence comprehension." It makes no claim about percent signal changes during sentence comprehension. 
• There's a reference to a paper "Retrieval and Unification of Syntactic Structure in Sentence Comprehension: an fMRI Study Using Word-Category Ambiguity." It's behind a paywall, and its abstract makes no claim about percent signal changes. 
• There's a reference to a paper "The cortical language circuit: from auditory perception to sentence comprehension."  It's behind a paywall, and its abstract makes no claim about percent signal changes. 
• There's a reference to a paper "fMRI reveals language-specific predictive coding during naturalistic sentence comprehension." The paper does not make any claim about a percent signal change during sentence comprehension. 
• There's a reference to the paper "Sentence complexity and input modality effects in sentence comprehension: an fMRI study." It's a study involving only 20 subjects, and does not claim in the main body of its text to have detected any percent signal change of 1% of higher. But there is a footnote in which the authors say that after doing some dubious-sounding fiddling with the data they got some kind of 1% difference of some type. We cannot have much confidence in such claim, as it stated only in a footnote. 
• There's a reference to the paper "Form and Content: Dissociating Syntax and Semantics in Sentence Comprehension." It does not make a claim about a percent signal change. 
• There's a reference to a paper "Location of lesions in stroke patients with deficits in syntactic processing in sentence comprehension." We can ignore it, because the issue is how much normal brain signals change during sentence comprehension.
• There's a reference to a paper "Time course of semantic processes during sentence comprehension: an fMRI study." It does not make a claim about a percent signal change. 

Rather than relying on the references Fedorenko has given, we can follow the alternate approach of doing a Google image search using the phrase " 'percent signal change' + ' sentence comprehension' ." This gives us papers such as these:

• The paper here involving sentence comprehension indicates a percent signal change of less than 1 part in 200 (less than half of a percent). 
• The paper here ("Neural correlates of syntactic movement: converging evidence from two fMRI experiments") reports a relatively large percent signal change of about 1 part in 100 for people listening to sentences. But the study group sizes are so small (involving only 11 subjects for one experiments, and 10 subjects for another experiment) that the paper cannot be counted as good evidence for anything. 
• The paper here ("Neural correlate of the construction of sentence meaning") must be disregarded because it reported a "percent signal change" based on EEG readings rather than fMRI readings, because it used a sample size of only 6 patients, and also because of the unreliability of looking for percent signal changes in the brain waves of seizure-prone patients. 
• The paper here ("Top-down and bottom-up contributions to understanding sentences describing objects in motion")  used an equally poor study group size of 12 subjects, and found a percent signal change of .05 percent, 1 part in 2000. 
• The paper here ("Language processing in the occipital cortex of congenitally blind adults") used one too-small study group size of 9 blind adults and another possibly halfway-adequate study group size of 22 control subjects with normal vision. Its Experiment 1 involving language processing found a percent signal change of less than .1 (less than 1 part in 1000) for the 22 control subjects. Another experiment involve language processing found a percent signal change of less than .5 (less than 1 part in 200) for the 22 control subjects.
• The paper here ("Brain activity associated with selective attention, divided attention and distraction") finds about a 1% percent signal change in the auditory cortex during language processing. But the study group size is an unimpressive 15 subjects. This auditory cortex region of the brain is not part of the "language processing" area claimed by Fedorenko. In her paper she refers to some regions of the brain and says, "These areas are distinct from the language network as well as from general-purpose sensory and motor areas, such as the primary auditory or primary motor cortex," apparently indicating that what she thinks is a "language network" in the brain is something outside of the auditory cortex. 

Typically there is no test/retest reliability when doing these little studies involving only 15 or 20 or 25 subjects. Robust evidence would only come from a much larger study group size. The New Scientist story below tells us that thousands of participants are needed for studies like these:

The purple bars on the graph shown above are very probably incorrect. Scientific papers documenting brain scanning during language comprehension do not show any robust evidence of a 1% or 2% change in brain activity during sentence comprehension. And there is no sound evidential basis for Fedorenko's diagram claiming to show the location of five language areas of the brain. 

As discussed here, a recent study indicated that 40% of MRI signals do not even correspond to brain activity, which further undermines the knowledge boasts that  Fedorenko has made, and provides another reason for disbelieving in the accuracy of her brain activity maps.  We read this: "Researchers at the Technical University of Munich (TUM) and the Friedrich-Alexander-University Erlangen-Nuremberg (FAU) have found that an increased fMRI signal is associated with reduced brain activity in around 40% of cases. At the same time, they observed decreased fMRI signals in regions with elevated activity." This is the kind of finding that should make us disbelieve Fedorenko's fMRI-based claims of a brain "language network." 

Link

The puff piece article in Quanta has a "softball questions only" interview with Fedorenko in which she states this:

"There’s a core set of areas in adult brains that acts as an interconnected system for computing linguistic structure. They store the mappings between words and meanings, and rules for how to put words together. When you learn a language, that’s what you learn: You learn these mappings and the rules. And that allows us to use this 'code' in incredibly flexible ways."

There is no robust scientific basis for these imaginative claims. To the contrary, the most powerful microscopes have examined very much brain tissue from very many still-living people and very many very recently deceased people, and no one has found the slightest trace of any sentence, word or letter stored in a brain. 

Later on Fedorenko states, " Then the language network parses that, finding familiar chunks in the utterance and using them as pointers to stored representations of meaning." Excluding the merely genetic chemical representations in DNA, there is zero evidence for any such "stored representations of meaning" anywhere in the brain.  Later she kind of gives away the speculative nature of her claims, by saying "there may well be cells that respond to particular aspects of language."  Note the use of "there may well be" rather than "there are."   

Fedorenko offers as evidence for this claim the preprint "Modality-Specific and Amodal Language Processing by Single Neurons."  It is a "Jesus in my toast" exercise in noise-mining. Recordings of neuron firings were taken after 1400 invasive microwire electrodes had been implanted in the brains of 21 very sick patients with the worst types of epilepsy. The authors tried to find some neurons that fired more often when certain words were spoken. Because neurons fire between 1 and 200 times per second, any such exercise will always be able to find some neurons that fired more often when certain were spoken. It is misleading to call such a thing "selectivity" as the paper did. Similarly, if I set up a computer program to try to correlate the passing of cars outside my house and the speaking of words on my TV set, and let the program run long enough, I will probably be able to find that certain words were spoken more commonly on my TV set when a car passed in front of my house.  That would be mere silly noise-mining, and that is all that is going on in this preprint, which fails to provide any decent evidence of any neuron responses to language different from what we would expect from chance variations, giving 1400 microwires implanted in brains. 

Noise-mining studies like this raise grave moral concerns about the reckless research endangerment of very sick epilepsy patients.  The sickest of epilepsy patients often have electrodes implanted for evaluation of where surgery should be done to reduce their symptoms. But such electrodes are usually not the microwire electrodes used for experiments like this. The deep implanting of such microwire electrodes in brains (70 microwires per patient for this study) has serious risks that are not medically justified, such as a risk of brain bleeding.  We should doubt any claim to have got any adequate degree of "informed consent" from the seizure-racked people who have these kind of deep microwire electrode implants. 

A scientific paper tells us, "Sixty-five years after single units were first recorded in the human brain, there remain no established clinical indications for microelectrode recordings in the presurgical evaluation of patients with epilepsy (Cash and Hochberg, 2015)." In other words, there is no medical justification for implanting microelectrodes or microwires in the brains of epilepsy patients. Complications from the insertions of such electrodes may include death, with two deaths reported here. I consider the exploitation and endangerment of sickest epilepsy patients in these type of poorly designed noise-mining experiments to be a serious moral scandal. 

We should laugh at Fedorenko's answer when she is asked whether there is an LLM (Large-Language Model) inside the brain, and she replies "pretty much." Nothing like that exists in the brain. An LLM or Large Language Model is a gigantic structure of data, computer programming software and data processing software built after some big server farm (consisting of very many individual computers) crunches data gobbled up by crawling the Internet, scooping up the contents of millions of web pages. That does not correspond to anything in a human brain. 

The idea of a language-parsing ability in a brain seems implausible when you consider that human brain anatomy has not changed in more than 20,000 years, but the main human languages (such as English) are less than 3000 years old. 

It is commonly claimed that the left half of the brain is needed for language. But if you read my series of posts labeled "loss of left half of brain," using the link here, you will find many cases that defy such a dogma.  For example, you will read of a case described like this: "He exhibited no lack of intelligence, yet after his death it was discovered that his left brain was practically destroyed and replaced by a watery substance." And you will read of how Beth Usher could still tell all of her "knock-knock" jokes just after the left  half of her brain had been removed to treat intractable seizures. You will read that on page 109 of a paper we read of three other cases of the removal of the left half of the brain, and we read that "speech and verbal comprehension were present immediately after left hemispherectomy in all three cases." The same post describes good verbal comprehension in patient E.C. after the left half of his brain was surgically removed. 

In a scientific paper ("Why Would You Remove Half a Brain? The Outcome of 58 Children After Hemispherectomy −−The Johns Hopkins Experience: 1968 to 1996") we read about how surgeons at Johns Hopkins Medical School performed fifty-eight hemispherectomy operations on children over a thirty-year period. Eleven of these children had the left hemisphere of their brains removed; most of the rest had the right hemisphere of their brains removed.  The paper states this:

"Despite removal of one hemisphere  [i.e. one half of the brain], the intellect of all but one of the children seems either unchanged or improved....Although there have been major concerns about loss of language after left hemispherectomy, all eleven of these children have regained virtually normal language....It is tempting to speculate, that the continuous electrical activity of these severely dysfunctional hemispheres interferes with the function of the other, more normal hemisphere. This might explain why motor function improves after hemispherectomy and why language recovers after removal of the dysfunctional left hemisphere, but does not seem to fully transfer before surgery. Perhaps it also partially explains intellectual improvement in these children after removal of half of the cortex. We are awed by the apparent retention of memory after removal of half of the brain, either half, and by the retention of the child’s personality and sense of humor."