You probably have a general understanding of the human brain: a network of nerve cells connected by synapses. Complex or abstract ideas emerge as a result of the firing of many of these nerve cells, or neurons. This means that a concept or memory or idea is the result of a distributed pattern of neural activity. In contrast to computer memory, which always follows the same pattern of 1s and 0s, it's more like networks in our brain weave a new tapestry every time we think about something. This understanding is common to most of us lay people who know a little about neurology and the brain, or who are interested in AI and the attempts to replicate human intelligence. 

Unfortunately, you may be badly out of date. As it turns out, in human brains, we also have a specialized type of cell called a concept neuron, which does what was long thought to be impossible: each of these cells encodes entire concepts, so that that single neuron fires whenever you’re exposed to a stimulus relating to that concept, or even when you think about it without an external stimulus. This would be like having a single neuron that fires when you see a photograph of your grandmother, hear her voice, read her name or perhaps even smell her familiar perfume. This single neuron thus has semantic invariance for the concept of your grandmother. This means that it fires whenever your grandmother is the topic of thought, regardless of the context or medium or the sense that stimulates your thought of her.

Dr. Florian Mormann, a physician and researcher who heads a working group on cognitive and clinical neurophysiology at the University of Bonn, told Salon in a video interview that “textbooks of neuroscience that still exist today often mention the grandmother neuron as an example of something you would never find in a brain. Because clearly, it would be so much more efficient to simply have a network of eight neurons if it can do the same job as 70 different neurons. So it seemed a no-brainer for everyone that there shouldn’t be grandmother neurons in any brain”. 

And yet, the no-brainer is seemingly wrong about the brain. As it turns out, grandmother neurons, or concept cells as they’re now known (or Jennifer Aniston neurons as they were called for a bit, as we’ll see) have been right there in our brains all along. And some scientists have known they are there, gradually learning more about them and their implications, over the past 20 years. It’s not a conspiracy: the news just hasn’t really trickled out to the rest of us. That’s partly because most scientists who physically get right inside skulls to study the brains inside them do so with the brains of non-human animals, so most brain research we hear about still doesn’t involve these neurons, which it seems exist exclusively in the human brain. 

Dr. Rodrigo Quian Quiroga, the director of the Centre for Systems Neuroscience at the University of Leicester, discovered concept cells twenty years ago. In research published in January, a team led by Quian Quiroga showed for the first time the way they respond to a given concept regardless of the context. This is in contrast to everything that is known about how memories are encoded in non-human animals, and it suggests that this might be a key adaptation behind human intelligence.

It’s not ethical to carry out invasive procedures on human brains. For this reason, we don’t often conduct single neuron recordings, a type of research that requires access right inside the skull, on humans. 

“So in the past 50 years [single neuron recordings have] been done on a huge scale in especially rodents and also in monkeys, plus a few other mammals. But it’s mainly those two species,” Mormann, who, like Quian Quiroga, is one of very few researchers to conduct single neuron studies on humans, told Salon. 

“And of course, the multi-million dollar industry of rodent research does not exist because we’re all so fascinated by what the rodent brain can do. Not at all. It is because we believe it could serve as a valid model of the human brain for human episodic memory,” Mormann added.

Instead, neuroscience has relied on the data gathered from doing such procedures on mouse and primate models. The assumption that underlies this is that a mouse or monkey brain serves as an adequate, if simplified, proxy for the human. And indeed, that’s been good enough to fuel decades of brain science. 

Single neuron recordings in humans 

Single neuron recordings are otherwise only done in rodent or primate models because brain surgery is never without risks, and those are not risks we usually ask humans to take without extremely good reason. A diagnostic procedure to ensure the surgeons don’t resect the wrong hippocampus when considering surgery for patients with epilepsy is something most of us would consider an extremely good reason.

So Mormann’s team, and a small number of other research groups, study the firing of individual neurons by inserting fine microwires inside the hollow tube of the depth electrodes that have to be implanted anyway in order to diagnose the location of a patient's seizures in hopes of curing their epilepsy. 

Two decades ago, shortly after single neuron recordings started to be used by scientists to take advantage of this rare opportunity to engage in ethical neurological research in human beings, Dr. Rodrigo Quian Quiroga, who now heads up the Neural Mechanisms of Perception and Memory Research Group at the Hospital del Mar Research Institute in Barcelona, showed that individual neurons could be extremely selective, with a single neuron firing predictably in response to images of a single animal, individual or place. 

---

Want more health and science stories in your inbox? Subscribe to Salon's weekly newsletter Lab Notes.

---

“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,” Quian Quiroga told Salon in a video interview. "And as I found this one, I found many others later on, and basically 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. It can be a person, it can be an object, it can be a place — whatever is relevant to the person or to the patient."

In that first work, described in 2005, UCLA neurosurgeon Itzhak Fried, Quian Quiroga, then his mentee, and other colleagues showed that a subset of neurons in the hippocampal formation, and more generally, in an area of the brain called the medial temporal lobe (MTL) that includes the hippocampus, fire in response to the subjects being shown pictures of a particular person under strikingly different conditions: at different ages, poses or contexts. A given neuron would even fire in response to reading the person’s name. Confusion is possible: for example, a Jennifer Aniston neuron might fire if you show the subject a picture of Lisa Kudrow, who played Phoebe on "Friends" alongside Aniston’s Rachel.

“Nobody expected that this neuron type could exist but we show very clearly that they do exist,” Quian Quiroga said. They dubbed these brain cells Jennifer Aniston neurons.

Whatever you call them, these neurons have not yet been found in rats, nor in monkeys, nor in any other mammal or organism — only in humans, despite Quian Quiroga challenging anyone, anywhere to find another creature with concept cells. But in the last two decades, no one has found them in anyone but humans, where they are reliably easy to find by testing neurons at random in the MTL.

A brief history of concept cells

Quian Quiroga’s team and others, like Mormann’s, conducted further studies that gradually revealed different properties of concept neurons in the MTL, primarily in the hippocampus, and how they behave in different conditions. Semantic concept neurons have also been found in different parts of the MTL, such as the amygdala and the entorhinal cortex. Recently, Mormann and his team identified smell-specific concept cells in the piriform cortex, which they dubbed olfactory concept cells. In fact, of 1,856 neurons tested in the piriform cortex in that study, 66 responded to both images and odors, meaning that they had semantic invariance with respect to the concept of the odor in question, firing in response to an image associated with the smell as well as the smell itself. 

Concepts encoded by concept neurons can be an animal, an article of clothing, a place, or a person. The concept can be evoked, and the associated neuron made to fire, by a direct stimulus such as seeing the person or hearing their voice; or without a stimulus, by imagery, free recall, or comparisons that reference the concept. 

“We show explicitly that [concept cells are] involved in forming and storing memories,” Quian Quiroga explained. “We had a series of experiments [where] we showed that these neurons are involved in the forming of memories, and they can do this very quickly. I mean, in just one shot, they can change the way they respond, and with this, they are encoding new experiences.”

He further argued that concept cells were fundamental components, or building blocks, of declarative memory. It’s not surprising, he felt, that we have neurons responding to concepts in an area of the brain associated with memory — after all, we do tend to remember concepts and forget details. The human way of remembering is very abstract. Most of us don’t remember precisely what a person looks like, what they are wearing, or the words they say in a conversation; rather, we focus more on the basic ideas. 

As the authors (including Mormann) of a 2020 study of how single neurons in the MTL are able to code abstract meaning write, “Although semantic abstraction is efficient and may facilitate generalization of knowledge to novel situations, it comes at the cost of a loss of detail and may be central to the generation of false memories.”

Still, that cost may well be the price of human intelligence.

“I started arguing,” Quian Quiroga recalled, “that this is a trait of human intelligence, and … one of the key aspects that distinguishes us from other animals: the fact that we just don’t focus on details, but we’re able to extract what is the important information and focus on that. And this is the way we store our memories, and this is the way we think so we are not bombarded with details.”

For humans, unlike our fellow animals, we just want the key point, the essential abstraction. That, Quian Quiroga maintains, is the way we remember, and the way we think. And that’s what he shows with the new study, published in Cell Reports. The research team recorded the activity of individual neurons while patients learned and then recalled two stories that described different situations but featured the same character or place. Nearly all of the neurons that fired initially did so without regard to the context, such that, as the authors explain, “taking all neurons together it is possible to decode the person/place being depicted in each story, but not the particular story.” 

The brain cells that fire during learning and memory are firing in response to the concept of that character and will fire in any context in which that concept features. Quian Quiroga believes that the development of language involved adaptation of neurons, common to all mammals, to this specialized purpose. Still, we don't know exactly when this arose, but perhaps it evolved gradually, Quian Quiroga theorized.

"I think in the last 100,000 years, the moment that [Homo] sapiens started uttering words and attributing meaning to things in terms of words,” Quian Quiroga said. "Then the sapiens started thinking in terms of words instead of pictures. I think that created the big phase transition of intelligence, the fact that we started thinking in terms of words. I think this created concept cells, because once you attribute the word into something then you get completely rid of the details. And that’s exactly what concept cells do."

A new frontier in neuroscience

“The striking thing that also since 2005 has gone largely ignored [by] rodent electrophysiologists is that this degree of semantic invariance, plus also context independence, is something we observe only in the human. There’s no other species in the animal kingdom where this has been convincingly reported, and even in humans, we only find them in the medial temporal lobe and not any other brain region,” Mormann explained. “To me, [it’s] one of the most seminal discoveries of the last 50 years, at least, but has been largely ignored. And the reason why it’s been largely ignored is, in my opinion, because this type of electrophysiological research traditionally cannot be done in humans, for obvious ethical constraints.” 

But as we've seen, there is one very particular circumstance in which it’s necessary to carry out such invasive procedures. Patients with seizures as a result of epilepsy may require exploratory surgery — invasive seizure diagnostics, it’s called — to determine if they would be good candidates for a neurosurgical resection where the seizure-generating area of the brain would be removed, taking away the condition and any neurological or cognitive deficits associated with it. 

“Our job is to make sure that we found the seizure-generating area so that we can then provide the patient with three pieces of information. One, their chances of becoming permanently seizure-free if we resect that area of the brain, which reflects how certain we are that we’ve identified the epileptic focus. Two is what price they’ve got to pay, because there is often some residual function that might be gone once we remove [the epileptic tissue]. And the third one is the complication risk,” Mormann explained to Salon with evident care. The history of lobotomy is such that no one would want to be associated with reckless brain surgery.

For a small proportion of this group of epilepsy patients, less than 10%, recording seizures using scalp EEG is enough to provide the necessary information. But in cases where they cannot reach a conclusion on those risks and benefits non-invasively, his team offers the patients diagnostic surgery in which they implant electrodes and use them to record seizures exactly where they are happening. The implants may be in for a week or more.

“The area that is being implanted the most [or used to be] is the medial temporal lobe, simply because it’s very well-shielded from the outside on both sides, and also because that is the region that’s mandatory for episodic memory formation,” Mormann said.

Henry Molaison — known for decades only as H.M. — became one of the most famous patients in neuroscience after losing his ability to form memories entirely in 1953. That incident, so unfortunate for him and so interesting for our understanding of memory, occurred due to a bilateral resection to control epileptic seizures that had blighted his life since the age of 10, possibly resulting from a minor bicycle accident. To say he had a bilateral resection means that the surgeons removed these structures on both hemispheres of the brain. The surgery was successful at curing the epilepsy, but left him with the inability to form new memories. (Interestingly, attempts to replicate the effects of Molaison’s surgery on memory in monkeys were unsuccessful at first, revealing that humans and monkeys use different parts of the brain for learning certain tasks.)

So it’s safer to resect just one hemisphere. But if you choose the wrong one to remove, the patient will still have seizures, and now may have impaired memory as well.

“That is why … it has become customary to be careful not to resect the wrong hippocampus, because simply, there’s no second attempt,” Mormann explained. 

So, Mormann believes, the research in humans has been largely ignored. In his view, semantic concept cells in humans are thought of as “a fancy version of place cells” at best. Place cells are cells found, so far, only in rodents, that fire at certain spots as a mouse or rat moves through a linear track, indicating a cell specific to certain locations. (As Mormann’s own recent research has affirmed, humans do also have location-specific neurons that play similarly specialized roles in spatial awareness.)

“I think there’s a paradigm shift involved in all this,” said Quian Quiroga. While colleagues have been teaching his work to undergraduates for years now, “there might be some inertia not to take this paradigm shift because … we assume that the human brain is kind of like an extrapolated version of the workings of the animal brain."

He added that neuroscientists, have described certain principles in animal models, "assume that these principles will also apply to humans, although maybe with a bit of a higher complexity. And I think what we’re showing is that … we shouldn’t take this for granted.”

Not that this invalidates the use of animal models in neurology. But it means that, rather than expecting a rat’s brain or a monkey’s brain to tell us all about how things work in humans, the differences between their brains and ours may be what’s really of interest. In fact, Quian Quiroga is busy with experiments aimed at quantifying the differences between the way humans process information in the hippocampus or in the memory system in general compared to what’s been described over the last fifty years of memory research using other animal models.

“This paper we just published is the first that I expect to be a series of studies because this is just showing the tip of the iceberg,” Quian Quiroga said. 

What concept cells could mean for AI

Of course, all those years of focus on mouse and monkey brains as models for how we think have also informed our work in artificial intelligence. Might this explain why, impressive though it is, we have not yet replicated the way humans really think? It wasn't until 2020, after fifteen years of experimental evidence supporting the existence of concept cells, that three researchers writing in Scientific Reports set out a theoretical justification for the possibility of such structures existing, and in fact for the likelihood of the existence of such cells in the hippocampus. "Three fundamental conditions, fulfilled by the human brain, ensure high cognitive functionality of single cells," the authors write.

Till now though, it was the very different, prevailing understanding of the brain that has guided our development of artificial brains: that the coordinated action of countless neurons — a neural network — is what allows for the representation of abstract concepts. Although of course we do use a distributed neural network for many aspects of our cognition, now it seems that representing entire concepts using specific cells with semantic invariance might be a key difference between human intelligence and that of other animals.

Perhaps also of interest to those hoping to build artificial brains, what Quian Quiroga’s work suggests is a possible explanation for how it is that anatomically, the brain of a human and that of a chimpanzee is not all that different. The human’s is bigger, but not so much as to explain the very considerable difference in intelligence. 

“So my point is not that the human brain is different,” Quian Quiroga said. “It’s that the human brain must be working differently. It is just the fact that in the chimpanzee brain, you will have a visual stimulus going all the way from visual processing areas into your memory system. So you form memories based on pictures, based on images. In the human brain, the visual system comes to a point, and then you extract a meaning from it, and it’s only the meaning and not the stimulus itself that goes into the memory system”

Our anatomy has barely changed from our common ancestor with chimpanzees. But language, it seems, has drastically changed how we use it.