Friday, March 21, 2025

Claimed Evidence for "Concept Cells" Is Just Noise-Mining Nonsense

 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.  The articles at Quanta Magazine often contain misleading prose, groundless boasts or the most glaring falsehoods. I discuss some examples of such poor journalism in my posts here and  here and here and here and here.

The latest piece of nonsense in Quanta Magazine is an article trying to persuade us that scientists have discovered "concept cells" in the brain. No such thing has occurred. What is mainly going on is noise mining,  the morally dubious exploitation of very sick epilepsy patients, and scientists trying to get citations and attention by wrongly applying unjustified nicknames to cells. 

The research discussed works like this:

(1) Electrodes are implanted in the brains of very sick epilepsy patients requiring surgery for epilepsy, supposedly for the sake of surgical evaluation on where to perform the surgery (although we should suspect that additional electrodes are being implanted so that this type of noise-mining research can be done). 

(2) The patients are shown some visual stimuli, and EEG readings of their brain waves are taken while they see this stimuli. 

(3) The data is then analyzed, with searches made for particular neurons that fired more often when some particular image was seen.  Scientists then make a triumphal declaration that a "concept cell" was found, on the basis of the claim that some neuron was firing more frequently than we would expect when some visual stimulus was seen. 

This is noise-mining, like someone searching 1000 photos of clouds looking for one that looks like the ghost of an animal.  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 seeing different things will be able to find a few neurons that seemed to fire more often when something was seen. 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 images 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 image was displayed. But that would be mere noise-mining. 

It is never justified to speak of single-neuron "responses" to a concept or an experience.  A single neuron does not respond to something a  person sees or recalls or thinks about. A neuron fires continuously at a rate of about 1 time per second or more, with random variations. It is always misleading to try and suggest a stimulus and response relation between a neuron firing and something someone saw or thought of or recalled.  This is like tracking many flu-infected people who each cough hundreds of times a day, and boasting about having found some "concept coughs," claiming that some of the coughs are a "response" to an image a person sees on a TV.  

The main paper discussed is the late 2024 paper "Concept and location neurons in the human brain provide the 'what' and 'where' in memory formation." The paper does nothing to find any evidence of memories stored in brains. All that is going is noise-mining of 3681 neuron firings in 13 epilepsy patients who had electrodes implanted in their brains. 

We should have no trust in the statistical analysis done in the paper, which is largely dependent upon a large body of programming code that looks like it is poorly written, and is performing all kinds of obscure or arbitrary convolutions and manipulations of data. You can see the programming code here. The EEG data is being processed in many a strange way, and is being passed through all kinds of contortion processes including many  doubly-nested loops doing God-only-knows-what, as in this example:

for k=1:numXbins

    for j=1:numYbins

        n(k,j)= sum ( wavs(:,k) <= ybins(j)+ybinSize/2 & wavs(:,k) > ybins(j)-ybinSize/2);

    end

end

No peer reviewer could ever untangle the programmatic "witches' brew" that is going on in the programming code of this paper. 

torturing data until it confesses

 A look at the peer reviewer comments on the paper gives us some hints about the paper's defects. One peer-reviewer asks this:

"These recordings come from epilepsy patients. How were possible epilepsy-related confounds mitigated?"

Here's what this comment refers to: epilepsy patients scheduled for surgery have all kinds of weird brain wave anomalies cropping up in their brain waves as recorded by EEG devices.  The possibilities for getting false alarms from EEG readings from epilepsy patients is endless. The paper authors respond to this question in an unconvincing way, by claiming that they were using EEG analysis software that did something to reduce such a problem. It is not a convincing response. 

When we read calculations of p-values in papers like this, we typically get no detailed discussion of how such a p-value was computed. We should have little trust in the accuracy of the calculated p-value. In the case of this paper, one of the peer reviewers says that he did his own calculation, and got a p-value number drastically different from  one of the p-values published in the paper.  

brain wave noise mining
The game of "keep torturing the data until it confesses"

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 rate of any single neuron  during the observation of sights by subjects would produce any evidence that such a neuron encodes or recognizes or represents or is sensitive to any concept. The idea that you would get meaningful evidence of such a thing from analyzing the firing of only 3681  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. If an individual neuron encoded a concept, we would expect that an experiment such as this would have less than 1 chance in a million of success, since 3681 is less than a millionth of the total number of neurons in a brain. 

In any study in which subjects are moving muscles, you can never assume that some neuron has some connection to a concept because greater firing activity occurred when that concept was displayed or chosen.  It is well-known that muscle movements abundantly contaminate EEG readings that indicate how much neurons are firing. So in any study in which subjects are not perfectly immobile, you have no way of knowing whether some increased neuron firing is merely due to some increase or difference in muscle movement. In this study subjects were not perfectly immobile when the EEG readings occurred. They were instead performing activities with their hands. For example, we read, "The participant was asked to confirm every image location by tapping it within the presentation time window (1.5–3.5 s)."

The idea of a concept cell is nonsense. No one has any coherent idea of how a single cell could represent a concept or why a cell would have a tendency to respond more frequently when a viewer is exposed to just one concept. Since the  paper is so dependent upon black-box spaghetti code that is programmatically fooling around with their gathered data in many strange and tangled ways, no one should have any confidence that evidence of a "concept cell" was found. 

Below is a visual (Figure 2) from a scientific paper that did single-cell recordings of the firings of individual neurons in monkeys. Each one of the vertical bars represents a spike in the firing rate of a neuron. There is no clear relation between the firing rate and the stimulus presented to the monkey,  The "clumpy-bursty" example was one cherry-picked to show the strongest evidence of response to the stimulus.  A visual like this makes clear that spikes or blips in the firing rate of a neuron occur randomly many times a day. It is deceptive to pick out a case where one of these blips occurs when the stimulus occurred, and to call such a blip a "response" to the stimulus. But such a deception is what is occurring in papers claiming to have found "concept cells." 

blips in neuron firing rates

The paper discussed above ("Concept and location neurons in the human brain provide the 'what' and 'where' in memory formation")  gives us no assertion that the microelectrodes implanted in the brains of these very sick epilepsy patients were inserted only for medical reasons, to evaluate where they should have surgery. Since we have got no such assertion, we should suspect that one or more of the microelectrodes were unnecessarily implanted in the brains of very sick patients, for the sake of this poor-quality study based on nonsensical assumptions. Implantation of microelectrodes in the brain comes with very serious risks.  People requiring epilepsy surgery are very sick people who should not be put at higher risk for the sake of low-quality research such as this. 

The paper discussed above implanted many microelectrodes in epilepsy patients. The type of electrode normally used for surgical evaluation of epilepsy patients is a much larger type of electrode called a macroelectrode. 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 in the brains of epilepsy patients. The paper tells us that the microelectrodes were "inserted through the hollow clinical macro electrodes, and protruding from the tips by ~4 mm." This was medically unnecessary and potentially hazardous insertion of many wires into the brains of very sick patients.  We seem to have here some very sick patients being put at needless risk merely so that junk science can be produced. 

reckless neuroscientist

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:

  1. Disruption of the blood–brain barrier (BBB);
  2. Tissue deformation;
  3. Scarring of the brain tissue around the implant, i.e., gliosis 
  4. Chronic inflammation after microelectrode implantation;
  5. Neuronal cells loss."
I strongly advise any people who participated in any brain scanning experiment or any neuroscience experiment involving electrode implants to permanently keep very careful records of their participation, to find out and write down the name of the scientific paper corresponding to the study, to write down and keep the names of any scientists or helpers they were involved with, to permanently keep a copy of any forms they signed, and to keep a very careful log of any health problems they ever have. Such information may be useful should such a person decide to file a lawsuit or a claim seeking monetary damages. 

Postscript: The term "gnostic cells" has sometimes been used to mean the same thing as "concept cells."

More noise-mining nonsense is found in the recent paper "Lack of context modulation in human single neuron responses in the medial temporal lobe," which you can read here. This time the study group is even smaller, consisting of only 9 very sick patients.  The authors have failed to make their code publicly available through any easy access method, as if they were embarrassed by their programming. But using the link here you can download their code as a .zip file, scan the .zip file for viruses, and then extract it, which is a laborious way to have to inspect code. I did that, and found the usual very-poorly-documented spaghetti code programming horror that you typically find in projects like this. We have undocumented loops like the one below, doing some unfathomable rigmarole contortions of the brain wave data:

for iresp=1:size(resp,2)
    if ~isempty(resp(iresp).trecall_phasic_ms)
        resp_rec = resp_rec + 1;
        active_cluster = resp(iresp).spike_times_Rec;
        phasic_recall = resp(iresp).trecall_phasic_ms;
        ntrials(1) = size(active_cluster{resp(iresp).responsive_Storiesindex(1)},1);
        ntrials(2) = size(active_cluster{resp(iresp).responsive_Storiesindex(2)},1);
        strength_pair = NaN*ones(max(ntrials),2);
    
        equiv_strength(resp_rec).chan = resp(iresp).channel_number;
        equiv_strength(resp_rec).class = resp(iresp).cluster;
        equiv_strength(resp_rec).pair = resp(iresp).responsive_Storiesindex;
            
        for istim=1:2 
            phas_vec = phasic_recall{resp(iresp).responsive_Storiesindex(istim)};
            spikes1 = active_cluster{resp(iresp).responsive_Storiesindex(istim)};
            spikes1 = arrayfun(@(k) spikes1{k}-phas_vec(k),[1:length(phas_vec)]','UniformOutput',false);
            strength_pair(1:ntrials(istim),istim) = cell2mat(cellfun(@(x) sum((x< twin_off) & (x> twin_on)),spikes1,'UniformOutput',0));
        end
            
    
        nsamp = min(ntrials);
        equiv_strength(resp_rec).samples = ntrials;
        equiv_strength(resp_rec).meandiff_Hz = (diff(nanmean(strength_pair)))/((twin_off-twin_on)/1000);
        equiv_strength(resp_rec).delta = sqrt(2/nsamp*(norminv(alpha)+norminv(alpha/2))^2);
        [equiv_strength(resp_rec).test_resu, equiv_strength(resp_rec).pval, equiv_strength(resp_rec).pooledSD, equiv_strength(resp_rec).meandiff] = TOST_2023(strength_pair(~isnan(strength_pair(:,1)),1), strength_pair(~isnan(strength_pair(:,2)),2), 'welch',equiv_strength(resp_rec).delta,alpha);
    end
end

And we have the equally ugly unfathomable bit of monkey business below:

for ii=1:length(spikes)
        if iscell(spikes{ii})
            all_spks=cell2mat(spikes{ii}');
        else
            all_spks=spikes{ii}(:);
        end

        spikes_tot=all_spks(all_spks < tmax_epoch+half_ancho_gauss & all_spks > tmin_epoch-half_ancho_gauss);
        spike_timeline = hist(spikes_tot,(tmin_epoch-half_ancho_gauss:sample_period:tmax_epoch+half_ancho_gauss))/ntrials(ii);
        n_spike_timeline = length(spike_timeline); %should be the same length as ejex
        integ_timeline_stim = conv(spike_timeline, int_window);
        integ_timeline_stim_cut = integ_timeline_stim(round(half_ancho_gauss/sample_period)+1:n_spike_timeline+round(half_ancho_gauss/sample_period));
        aver_fr{ii} =  smooth(integ_timeline_stim_cut(which_times),smooth_bin);

        % subplot(length(spikes),1,ii)
        % plot(ejex(which_times(1:downs:end)),aver_fr{ii}(1:downs:end))
        % maxi(ii) = max(aver_fr{ii});
    end  

I can merely say the more time a programmer spends looking at the programming code used by this paper, the less confidence he will have that the paper did anything to establish the existence of "concept cells." This is black box "witches' brew" monkeying around with brain wave data that is best described with the phrase "they kept torturing the data until it confessed in the weakest whisper."

A visual from the paper gives you the type of data that was being tortured to try to get something. Random cases of more noise blips from noisy and extremely variable neurons (during a few seconds) are being passed off as "concept responses." It's like someone tracking each and every noise blip from a nearby street construction crew using a jackhammer,  and trying to correlate particular noise spikes with images appearing on his TV set. That would be a very silly case of noise mining, and what is going on in this paper is just as silly. 


One of the authors of this poor-quality paper is Rodrigo Quian Quiroga, who has long been quoted as making a misleading claim about a "Jennifer Aniston" concept cell.  For example:
  • In a 2017 article "Concept cells: the building blocks of declarative memory functions" by  Quian Quiroga, he incorrectly stated, "One of the first such neurons found in the hippocampus fired to seven different pictures of the actress Jennifer Aniston and not to 80 other pictures of known and unknown people, animals and places." This was a claim that did not match the data in his 2005 paper where we have in Figure 1A a visual depiction of more than 50 firings of that neuron when the subject was shown a picture other than Jennifer Aniston.
  • In Figure 1 of the year 2020 "Searching for the neural correlates of human intelligence" article by Quian Quiroga, he misleadingly states that the Jennifer Aniston neuron "did not respond to about 80 pictures of other persons," a claim which does not match the data in his 2005 paper where we have in Figure 1A a visual depiction of more than 50 firings of that neuron when the subject was shown a picture other than Jennifer Aniston.  
  • In a 2025 Salon article with many misstatements,  Quian Quiroga made this incorrect statement: "'Twenty years ago … I was doing experiments with a patient, and then I showed many pictures of Jennifer Aniston, and I found a neuron that responded only to her and to nothing else...It was very clear that in an area called the hippocampus that is known to be critical for memory, we have neurons that represent, in this case, specific people, or in general, specific concepts. "  The claim does not match what was reported in Quian Quiroga's 2005 paper, where  we have in Figure 1A a visual depiction of more than 50 firings of that neuron when the subject was shown a picture other than Jennifer Aniston.  
 In fact, Figure 1A of the 2005 paper shows that the famed "Jennifer Aniston neuron" did fire more than seven times when a picture of a basketball player was shown, that it did fire more than seven times when a picture of another basketball player was shown, that it did fire more than five times when a picture of a snake was shown, and that it did fire more than six times when a picture of the Leaning Tower of Pisa was shown. The paper says this: "To hold their attention, patients had to perform a simple task during all sessions (indicating with a key press whether a human face was present in the image)."  It is known that the muscle movements can abundantly affect EEG readings and whether a neuron fires. A difference between key presses when the Jennifer Aniston picture was shown and things other than faces were shown can easily account for why one neuron may have fired differently when pictures of Jennifer were shown, without any need at all to evoke an idea of a "concept cell" for Jennifer Aniston. Maybe Quian Quiroga's misstatements on his "Jennifer Aniston neuron" had something to do with the fact that he later managed to get a book deal, a deal for a book with a title mentioning his claimed "Jennifer Aniston" neuron. 

When he discusses this "Jennifer Aniston neuron" outside of the original paper, Quian Quiroga seems to never mention that the subject with that neuron was asked to press a button whenever he saw a face, and that muscle movements are known to increase brain wave spikes picked up by electrodes or EEG devices.  That seems like the "secret sauce" behind his "Jennifer Aniston" neuron. 

There is nothing very impressive about this case of the neuron's firing more often while someone saw a few pictures of Jennifer Aniston. While unlikely to occur on any one day, it is the kind of result you would expect an eager noise miner to produce after spending very long periods of time looking for some result better than chance in a dataset of random data. Similarly, if someone is very eager to find a cloud shape looking like the ghost of an animal, and he spends day after day scanning photos of clouds looking for such a thing, he will probably find one or two clouds that look rather like animal ghosts. Activity like this is correctly described as misguided noise-mining "Jesus in my toast" pareidolia, aided by misstatements in which the meager result is described in a misleading way. 

No comments:

Post a Comment