How We Got Misinformed About "Grandmother Cells"

Each neuron fires between about 1 and 200 times per second, with the firing rate being unpredictable. So neurons are a noisy, unpredictable signal source; and that kind of source provides opportunities for noise mining and pareidolia, the occasional finding of some desired pattern by people scanning noisy, variable data looking for such a pattern. Similarly, at a restaurant that makes 200 pieces of toast every day using different types of bread, there is an opportunity for noise mining, in which someone checking each piece of toast may eventually claim to see the face of Jesus in a slice of toast. 

Let us look at the history of claims of "grandmother cells." The term refer to some neuron that might allegedly respond only when a person sees some particular type of visual, such as a picture of the person's grandmother. The 2002 article "Genealogy of the 'Grandmother Cell'" by the late Charles G. Gross gives us some background on how the idea of such a cell got started. Gross tells us this:

"The term originated in a parable Jerry Lettvin told in 1967. A similar concept had been systematically developed a few years earlier by Jerzy Konorski who called such cells 'gnostic' units."

So according to Gross, the concept of a "grandmother cell" arose independently of observations, without any empirical warrant.  But then Gross starts telling an unwarranted self-serving tale that evidence was found supportive of such an idea. He claims, "In the early 1970s, my colleagues and I working at M.I.T. in Cambridge, Massachusetts, reported visual neurons in the inferior temporal cortex of the monkey that fired selectively to hands and faces (Gross and others 1969, 1972; Gross 1998a)." Gross is here engaging in self-citation. Let us look at the papers Gross refers to, and see whether they actually gave any evidence to back up such a claim. 

• The 1969 paper "Visual Receptive Fields of Neurons in Inferotemporal Cortex of the Monkey" by Gross and others which you can read here. We have no specific data backing up any claim that anything had been found like a neuron that only responds to some particular image. We merely have this vague statement: "There were several units that responded most strongly to more complicated figures. For example, one unit that responded to dark rectangles responded much more strongly to a cutout of a monkey hand, and the more the stimulus looked like a hand, the more strongly the unit responded to it." The paper gives no data backing up such a claim. 
• The 1972 paper by Gross is the paper "Visual properties of neurons in inferotemporal cortex of the macaque." Only the first page of the paper is publicly available here. That page makes no claim backing up claims of anything like a grandmother cell. 
• The "Gross 1998a" citation is a citation of the book "Brain, Vision, Memory" by Gross, which you can read here.  On page 198 Gross claims that "he did not publish a full account of a face-selective neuron until 1982," which shows that the previous two citations were inappropriate. On the same page Gross misspoke by claiming that "soon thereafter, a flood of papers on such cells appeared." No such flood occurred. He mentions a 1982 paper by Perrett, Rolls and Cann ("Visual Neurones Responsive to Faces in the Monkey Temporal Cortex"), which you can read here. 

Nothing that is in any of these citations supports the claim that anything like a face-selective cell or a hand-selective cell was discovered. If we look at the 1982 paper by Perrett, Rolls and Cann ("Visual Neurones Responsive to Faces in the Monkey Temporal Cortex"), which you can read here, we also find nothing impressive. The paper claims that "Of the 497 cells recorded in the STS region there was a sub-population of at least 48 cells which gave responses to the sight of faces that were two to ten times as large as the responses to other stimuli tested." There is no claim that these cells fired only when other faces were shown, and Figure 3 (cherry-picked as the strongest evidence of a "face responsive cell") shows the cell firing many times when things other than faces were shown. The graphs in the diagram are examples of cherry-picking, showing results from a few cells that seemed to fire the most when the subject was shown faces. 

Some mathematical analysis will show how unimpressive the result discussed above. In the study there were five types of sensory stimuli: faces, gratings/geometric stimuli, complex 3D stimuli, somatosensory stimuli, and auditory stimuli. Let us imagine that we are recording how 497 cells respond when a subject is exposed to one of a small number of categories of stimulus, such as five.  Given a high random variability in how the cells respond, with the firing rates varying randomly between 1 and 200 times per second, and given a relatively small number of trials, and only a small number of types of stimulus (such as only five), we would expect that by chance there would be about 10% of these cells that would fire twice as often or more when a subject is exposed to one of the five types of stimulus. So the reported result that "there was a sub-population of at least 48 cells which gave responses to the sight of faces that were two to ten times as large as the responses to other stimuli tested" is not something unexpected, assuming purely chance results, and no actual "face sensitivity" or "face selectivity" going on in the cells. 

The claim by Gross to have discovered neurons that "fired selectively to faces and hands" was false. Neither he nor anyone else discovered any such thing. All that was going on was noise-mining.  Monkeys were being shown different visual stimuli, including faces and things that were not faces. The firing of hundreds of neurons were recorded, and researchers were drawing attention to the cells that happened to have the highest firing rate when faces were shown. No evidence was being presented of more neuron firing during face observation than we would expect to see from a random set of randomly firing cells that fired with a high variability. 

Later in the 2002 article "Genealogy of the 'Grandmother Cell'" by Charles G. Gross, Gross makes this claim: "Starting 10 years later, these finding were replicated and extended in a number of laboratories (e.g., Perrett and others 1982; Rolls 1984; Yamane and others 1988) and were often viewed as evidence for grandmother cells." The references do not actually refer to any papers providing evidence for grandmother cells. The 1982 Perrett paper is discussed above, and did not find any such evidence, but merely claimed "Of the 497 cells recorded in the STS region there was a sub-population of at least 48 cells which gave responses to the sight of faces that were two to ten times as large as the responses to other stimuli tested."  The Rolls 1984 paper is the paper "Neurons in the cortex of the temporal lobe and in the amygdala of the monkey with responses selective for faces."  It is merely another paper picking out some cells out of hundreds that fired more often when faces were shown, while also firing when things other than faces were shown. 

None of the papers that Gross has cited could intelligently be interpreted as evidence for grandmother cells, so Gross misleads us badly by claiming that such papers "were often viewed as evidence for grandmother cells." Later Gross confesses, "However, most of the reported face-selective cells do not really fit a very strict criteria of grandmother/ gnostic cells in representing a specific percept, that is, a cell narrowly selective for one face and only one face across transformations of size, orientation, and color (Desimone 1991; Gross 1992)." At the end of the paper, Gross deceives us by trying to make it sound like these alleged "face selective" cells may be something like "grandmother cells." But no evidence he has presented or cited has given any evidence for such "grandmother cells." 

The next big development on this topic occurred when scientists started reading the firings of neurons in individual humans. This is something that cannot be done by simply having a person wear an EEG cap on his head. The reading of firings of individual neurons in humans requires the implanting of electrodes into the brain.  Some people with drug-resistant epilepsy have electrodes implanted in their brains so that doctors can figure out where is the best place to do surgery to help cure their epilepsy. Neuroscientists have tried to leverage the implanting of such electrodes, to study the firing of individual neurons in the human brain. 

This has often been a morally objectionable type of activity by neuroscientists. The type of electrodes implanted in a brain to evaluate a patient for epilepsy are called macroelectrodes.  The type of electrodes implanted to record the firing of individual neurons are called microelectrodes. There is never any medical justification for implanting microelectrodes in addition to macroelectrodes. 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. 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."

What is going on with attempts to find something like grandmother cells in humans is typically a morally objectionable affair in which very sick people are being put to unnecessary risks for the sake of scientists seeking fame and glory. Such affairs are so morally dubious that we should have a natural tendency to distrust the statements of scientists doing such research, just as we should have a natural tendency to distrust the statements of any person engaged in a reckless or shady activity. 

Similar to claims of a "grandmother cell" are claims of a "Jennifer Aniston neuron" that was activated only when a epileptic subject was shown a picture of Jennifer Aniston. The claim is unfounded, and does not match the data in the original paper. For a discussion of the shady business that went when claims like this were made, see the last seven  paragraphs of my post here. 

At the link here, a vision scientist describes some of what is going in studies like the studies mentioned above:

"Neuroscience, as it is practiced today, is a pseudoscience, largely because it relies on post hoc correlation-fishing....As previously detailed, practitioners simply record some neural activity within a particular time frame; describe some events going on in the lab during the same time frame; then fish around for correlations between the events and the 'data' collected. Correlations, of course, will always be found. Even if, instead of neural recordings and 'stimuli' or 'tasks' we simply used two sets of random numbers, we would find correlations, simply due to chance. What’s more, the bigger the dataset, the more chance correlations we’ll turn out (Calude & Longo (2016)). So this type of exercise will always yield 'results;' and since all we’re called on to do is count and correlate, there’s no way we can fail. Maybe some of our correlations are 'true,' i.e. represent reliable associations; but we have no way of knowing; and in the case of complex systems, it’s extremely unlikely. It’s akin to flipping a coin a number of times, recording the results, and making fancy algorithms linking e.g. the third throw with the sixth, and hundredth, or describing some involved pattern between odd and even throws, etc. The possible constructs, or 'models' we could concoct are endless. But if you repeat the flips, your results will certainly be different, and your algorithms invalid...As Konrad Kording has admitted, practitioners get around the non-replication problem simply by avoiding doing replications.” 

Later in the same scientist's blog, we read this year 2023 comment: "Articles published during the past decade bemoaning the inability of mainstream neuroscience to generate replicable or even reproducible outcomes are too many to count."  In the same post the scientist states this:

"If we weren't living it, it would be hard to imagine how a research culture could have strayed so far from the path of rationality as has the culture of neuroscience. Fundamental problems in theory and method have long been flagged (e.g. Teller, 1984; Jonas & Kording, 2018; Brette, 2019), but critiques have left barely a trace on the hard-beaten track of routine, mainstream practice."