Every year the journal Nature publishes many junk science neuroscience papers guilty of very bad examples of Questionable Research Practices. It is a huge mistake for anyone to think that the most prestigious journals such as Cell and Nature only publish high-quality neuroscience research. When it comes to neuroscience research, the standards of such "big name" journals are very low. The journal Nature also regularly publishes false-headline articles about neuroscience research, articles with some of the most misleading clickbait boasts to be found anywhere. Let us look at the latest example of a groundless boast the journal Nature is peddling. It is the unfounded claim that someone discovered "neurons that encode words’ meaning." The idea is nonsensical. No one has any idea of how a neuron could possibly encode the meaning of a word.
In the Nature article peddling this boast, we have a subtitle of "For the first time, scientists identify individual brain cells linked to the linguistic essence of a word." Always distrust anyone claiming to have done something in the field of neuroscience "for the first time." What it generally means is some research making some boast that has not been reproduced by other scientists. We read a description of what was going on:
"But Williams and his colleagues found a unique opportunity to look at how individual neurons encode language in real time. His group recruited 10 people about to undergo surgery for epilepsy, each of whom had had electrodes implanted in their brains to determine the source of their seizures. The electrodes allowed the researchers to record activity from around 300 neurons in each person’s prefrontal cortex.
As participants listened to multiple short sentences containing a total of around 450 words, the scientists recorded which neurons fired and when. Williams says that around two or three distinct neurons lit up for each word, although he points out that the team recorded only the activity of a tiny fraction of the prefrontal cortex’s billions of neurons."
This is noise-mining. Neurons in the brain fire continuously, at a rate between 1 time per second and 200 times per second. Anyone tracking the firing of 300 neurons while someone is listening to speech will be able to find a few neurons that seemed to fire more often when something was said. Similarly, if someone records the ups and downs of 300 stocks on the New York Stock exchange, and tries to correlate them with the occurrence of words coming from his television set showing an old movie, he will be able to find a few stocks that went up or down more often when some particular word was spoken. But that would be mere noise-mining.
The diagram below is from the scientific paper here ("Neuronal firing rates diverge during REM and homogenize during non-REM"). The authors say that they sorted neuron rates into five quintiles. Evidently neurons fire in a fluctuating manner at a rate between about 1 time per second (the top line) and about 100 times per second (the bottom line). This is just the kind of noisy, fluctuating data that noise-miners love to have. Anyone analyzing data so random and fluctuating can find some neurons that seem to fire more often when compared to some unrelated factor such as a word someone is hearing.
The Nature article is promoting the paper "Semantic encoding during language comprehension at single-cell resolution." Looking at the paper, I quickly find much use of misleading language. We read this:
"We observed that many of the neurons responded selectively to specific word meanings. The selectivity or ‘tuning’ of neurons reflects the degree to which they respond to words denoting particular meanings (that is, words that belong to specific semantic domains)."
Never trust a physical scientist using the word "selectively" or "selection." We have a 160-year history of scientists using the word "selection" in a misleading way, which began with Charles Darwin coining the term "natural selection" to refer to a blind, mindless process that was not actually selection. "Selection" is a word meaning a choice by a conscious agent. The neurons referred to above were in no sense selecting anything. It was merely allegedly found that some of these neurons were firing more often when certain words were spoken. This is exactly what we would expect to find by chance, given 300 random neurons with random firings that were compared to the playing of 450 random words. It is no evidence at all of either encoding or selection going on in neurons. All that's going on here is noise-mining. You could find a similar result by having a computer program with voice recognition abilities compare words coming from one of your television sets with words spoken when some other TV in your home played some television series shown on another channel. After data collection and automatic programmatic analysis, it would be found that some type of words might be heard more often on the second channel when some particular word was spoken on the first channel. There would be no causal relation involved, and the result would be meaningless noise-mining.
To help facilitate their noise-mining, the authors of the paper have "greased the wheels" by introducing something they call "semantic domains." Since the authors were free to arbitrarily put words in whatever "semantic domains" they wished, this helped them to see what they were hoping to see. Throughout the paper, verbs and nouns are used in a misleading way. We hear of the "selection" and "selectivity" of neurons that were not selecting anything, and we hear the claim that "many neurons distinguished words from nonwords," although no evidence has been provided of any actual distinguishing or recognition. What's going on is that the authors are merely crunching the random data of neuron firings, and correlating some cases of more frequent firings with the simultaneous occurrence of spoken words. This is being passed off as neurons "distinguishing" something or "selecting" something. Similar nonsense would be going on if I tracked the random ups and downs of 300 stocks, correlating such data with 400 words coming out of my TV set while an old movie played, and then said that certain stocks on the stock exchange "selected" or "distinguished" particular words.
Our authors then move on to misusing the word "representations," a word that today's neuroscientists are constantly misusing, in cases where they have zero robust evidence that any such thing as a representation is occurring. No evidence has been provided here of brains encoding words or recognizing words or selecting words or distinguishing words or representing. All that has gone on is noise-mining.
A look at the programming code used for this experiment shows all kinds of arbitrary data fiddling going on, what smells like a byzantine rigmarole affair of "keep torturing the data until it confesses in the weakest voice." Opening up the files using the link above, an experienced programmer like me can sniff many a sign of hard-to-justify programmatic convolutions going on, with the data being very frequently passed through all kinds of weird loops that are doing God-only-knows-what. I very much doubt that any of the programmers of this code or any of the authors of the paper fully understand what was going on with these data manipulations. Graphs in the paper are based not on the original raw data gathered from brains, but the output after such data was programmatically manipulated in many a strange and arbitrary way.
This experiment was based on nonsensical assumptions. The brain has billions of neurons and trillions of synapses. There never was any reason to believe that studying the firing of any single neuron during the hearing of words by subjects would produce any evidence that such a neuron encodes or recognizes or represents or is sensitive to any word. The idea that you would get meaningful evidence of such a thing from analyzing the firing of a few hundred randomly selected neurons (from a brain of billions of neurons) never made any sense, particularly given the extremely large random variations in the firings of noisy neurons.
Normally, noise-mining is a harmless bit of nonsense. For example, someone may try to analyze the ups and downs of summer temperatures and windspeeds in coastal China and try to correlate them with the wins and losses of baseball teams in the United States, maybe finding one or two teams that seem to perform better when it is windy or warm in coastal China. That's harmless nonsense. But in this case the authors may not merely have engaged in harmless nonsense.
A medical page here describes how electrodes are implanted to detect where seizures come from. It says this:
"We are going to place several (~10-15) electrodes into your brain. These electrodes are thin, floppy wires about the thickness of a spaghetti noodle."
But what is described by the "Semantic encoding during language comprehension at single-cell resolution" paper I am discussing is something vastly more than what is mentioned above: the implanting of a very large number of microelectrodes capable of recording the activity of individual neurons. Below is only a small part of the description of how the subjects seemed to undergo additional electrode implanting, with more than 900 microelectrodes being implanted into the brain (the "cortical ribbon" refers to the cortex of the brain):
"For the silicon microelectrode recordings, sterile Neuropixels probes31 (version 1.0-S, IMEC, ethylene oxide sterilized by BioSeal) were advanced into the cortical ribbon with a manipulator connected to a ROSA ONE Brain (Zimmer Biomet) robotic arm. The probes (width: 70 µm, length: 10 mm, thickness: 100 µm) consisted of 960 contact sites (384 preselected recording channels) that were laid out in a chequerboard pattern."
The news article in Nature says, "His group recruited ten people about to undergo surgery for epilepsy, each of whom had had electrodes implanted in their brains to determine the source of their seizures." But the scientific paper makes no mention of epilepsy or seizures. Instead it tells us that "All participants included in the study were scheduled to undergo planned awake intraoperative neurophysiology and single-neuronal recordings for deep brain stimulation targeting." A table of the paper lists only patient diagnoses other than epilepsy: Parkinson's disease, essential tremor and cervical dystonia. Oops, it seems the journal Nature can't get its story straight.
Deep brain stimulation is sometimes used for Parkinson's disease, although it is a risky technique (the paper here reporting a 13% incidence of "serious adverse effects" with such a treatment). Deep-brain stimulation for Parkinson's typically involves targeting regions deep in the brain, regions other than the cortex: the thalamus, the subthalmic nucleus, and the globus pallidus. But the paper I am criticizing got data from cortex regions far away from such regions. It tells us, "Single-neuron recordings were obtained from the left language-dominant prefrontal cortex."
We are told that decisions about the electrode insertions "were made independently of any study consideration," but you may reasonably doubt such a claim if the unspecified persons making such decisions were the same people benefiting from the publication of the paper. In the medical world there is a massive history of doctors claiming without much credibility to have made treatment decisions or recommendations "independently" of some financial or professional benefit they may derive from such recommendations or decisions. For example, innumerable doctors claimed (without much credibility) to have recommended opioid prescriptions "independently" of financial gain they got from such recommendations. Similarly, doctors at clinics with profitable fancy scanners may claim to recommend expensive scans "independently" of the doctor's financial benefits from such scans (a claim we should doubt). And many doctors often recommend unnecessary surgery treatments profitable for them, and claim to have reached such recommendations "independently" of financial profits the doctors get from such treatment; but you should often be skeptical of such claims of decisions made independently of the doctor's financial benefits.
We may wonder if the ten subjects each getting hundreds of microelectrode implants underwent very significant risk largely for the sake of a badly conceived experiment that did nothing to provide robust evidence about neurons and cognition.
Here is a quote from a scientific paper:
"The effects of penetrating microelectrode implantation on brain tissues according to the literature data... are as follows:
- Disruption of the blood–brain barrier (BBB);
- Tissue deformation;
- Scarring of the brain tissue around the implant, i.e., gliosis
- Chronic inflammation after microelectrode implantation;
- Neuronal cells loss."
There is zero robust evidence of any kind of encoding going on in the brain other than the genetic coding that goes on DNA, which is merely low-level chemical coding in which amino acids are represented. Any claims to the contrary will not hold up to careful scrutiny, and will typically be found to be appeals to poorly designed experiments engaging in Questionable Research Practices, experiments not producing results that were well-replicated.
The structure of human neurons has not changed in many thousands of years. Conversely, words in a language have typically only been used for centuries. Words have meanings that often gradually change from decade to decade. For example, the English word "chip" has multiple meanings, mostly meanings that arose relatively recently. A person using the phrase "chip" may refer to a computer component, a plastic poker chip, or a potato chip, all human inventions of recent centuries. The English word "hack" now has a computer-related meaning it did not have when I was born, and the English word "bug" now has two meanings it did not have in 1930, one referring to a software defect and the other referring to electronic surveillance. The idea that neurons (with structures unchanged in 5000 years) encode human words (mostly less than thousands of years old and often with meanings only decades old) makes no sense. For more on this, see my post "Young Age of Languages Contradicts Claims of Neural Storage of Linguistic Information."
The encoding of words occurs in things such as dictionaries, which describe the multiple meanings of words. The encoding of a word is something far more complicated than merely a higher response rate. My eyebrows may raise if I am swimming and you shout, "Shark!" But that would not be a case of my eyebrows encoding the word "shark." And similarly, you would not show a neuron encoded a word if you gave data suggesting the neuron fired more often when the word was spoken.
Postscript: Another noise-mining paper ("Default mode network electrophysiological dynamics and causal role in creative thinking") fails to provide any robust evidence that brains produce creative thinking, but does provide more reasons for suspecting that patients are being put at risk for the sake of junk science papers. We have another paper using too-small study group sizes (13 patients total, with study group sizes such as 9 and 10 subjects). No blinding protocol was used, a necessity for any study like this to be taken seriously. The patients were intractable epilepsy patients due for brain surgery, but we are told that "the electrodes were selected based on anatomical and clinical considerations (not part of the suspected seizure onset zone)," We may wonder whether the patients had electrodes inserted other than what was necessary to determine where they should have surgery to help their seizures. We hear in the paper of brain-zapping as part of the experiment, which can produce seizures in epileptic patients. 15 subjects per study group is the minimum for any experimental study seeking correlation evidence, and for most effect sizes the required study group size is much higher (30 or more). A 2021 paper ("Increasing the statistical power of animal experiments with historical control data" by V. Bonapersona et. al. ) tells us, "For a common effect size of Hedge’s g= 0.5 (Welch’s independent samples t-test, α=0.05), ten animals per group would correspond to a statistical power of 18%, 30 animals per group to 48% power and 65 animals per group to 81% power." 80% statistical power is regarded as good, and 18% statistical power is very weak statistical power, with a high likelihood of a false alarm. Again we may sadly wonder how many very sick patients are put at risk for noise-mining junk science papers.
No comments:
Post a Comment