
Forget Network Layers—Cortical Columns Think Like Graphs
What this covers
Charles Simon presents a structural argument about how the brain stores and processes knowledge, grounded in neuroscience but aimed at redirecting AI design. The video proposes that the neocortex operates as a biological graph processor rather than as a statistical engine like modern neural networks. Its central claim is that cortical columns—the brain's repeating microarchitecture, each roughly 100 neurons—function as nodes in a graph, where relationships between concepts (is-a, has-a, part-of) are wired in from birth and refined through experience rather than learned through gradient descent. Simon argues that this graph-based approach solves fundamental problems: it handles inheritance naturally (Fido inherits the property "has fur" from being a dog) and accommodates exceptions without retraining (Stubby lacks a tail despite being a dog). The video sets this against the background that the neocortex is composed of millions of nearly identical columns that specialize only based on their input signals, suggesting a single repeated general-purpose module underlies all cortical function.
The substantive disagreement centers on how knowledge is best represented in silicon. Simon contends that artificial neural networks, by assuming smooth statistical gradients and uniform layers, cannot achieve reasoning, explain concepts, or handle exceptions the way a graph model does—a claim he frames as essential to building AI that genuinely understands rather than merely recognizes patterns. He also emphasizes efficiency: the brain runs on 20 watts and demonstrates understanding far beyond current systems, a difference he attributes to architectural principle rather than brute force. The video explores how a tiny number of AND-gates and pre-existing synaptic scaffolding can encode relationships, and how bidirectional links enable both forward retrieval (activate Fido, retrieve its attributes) and backward search (activate dog, find all instances). Charles Simon acknowledges this remains a fringe position within neuroscience and AI, but presents it as the necessary direction if the field wants to build systems that think rather than predict.
Charles Simon argues that the neocortex functions as a graph-based knowledge processor where cortical columns are nodes representing concepts, connected by precise relationships, which explains human understanding and reasoning capabilities that current AI systems lack.
- Cortical columns are uniform repeating structures that function as nodes in a knowledge graph, each containing ~100 neurons organized through six layers
- Graph-based symbolic reasoning with inheritance and exceptions enables meaningful conceptual understanding, something statistical neural networks cannot achieve
- The brain's structural uniformity combined with specialized function suggests a single repeated computational module that operates as a biological graph processor
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Knowledge can be represented as a graph of relationships with inheritance and exceptions, such that 'Fido is a dog' lets Fido inherit 'has fur' on the fly, while an exception (Stubby has no tail despite dogs having tails) overrides just that one attribute, enabling enormous compression of information.
“if a dog has fur, then phto inherits the has fur attribute on the fly. But if stubby is a dog and dogs have tails, yet Stubby has no tail, the exception overrides just that attribute.”
If we want to build AI that actually thinks, we need to stop layering up weights and start wiring in structure, because today's AI lacks understanding, generalization, and common sense precisely because it does not represent knowledge as a graph of concepts and relationships.
“If we want to build AI that actually thinks, we need to stop layering up weights and start wiring in structure.”
The brain runs on just 20 watts yet shows understanding and common sense far beyond today's AI systems.
“Your brain runs on just 20 watts, yet it shows understanding and common sense far beyond today's AI systems.”
The cortex is built from a single repeated module—a general-purpose building block—that serves as the fundamental computational unit across the brain.
“This suggests the cortex is built from a single repeated module, a generalurpose building block.”
Unlike computer chips where different regions look and function very differently, the neocortex is surprisingly uniform in appearance and organization across different regions.
“Unlike a computer chip where different regions look and act very differently. The Neoortex appears surprisingly uniform.”
The neocortex consists of millions of repeating structures called cortical columns that are not identical but remarkably similar, raising the question of how uniform structures produce vastly different capabilities from vision to language to reasoning.
“The Neoortex appears surprisingly uniform. Unlike a computer chip where different regions look and act very differently. The Neoortex consists of millions of repeating structures called cortical columns. These aren't identical, but they're close. Which raises a new question. If all these columns are so similar, how do they produce such vastly different capabilities from vision to language to reasoning to love?”
In early human brain development, all cortical columns are essentially identical, but they specialize based on the signals they receive—such as from the eyes, ears, or other cortical regions—so visual cortex becomes tuned to edges and colors while prefrontal cortex learns planning and reasoning, all starting from the same basic circuitry.
“In early human brain development, all columns are basically the same, but based on what signals they receive, say from the eyes, the ears or from other cortical regions, they specialize. So the visual cortex becomes tuned to edges and colors while the preffrontal cortex learns to plan and reason. But the key is they all start with the same circuitry.”
The brain works not as a statistical engine but as a biological graph processor where concepts are nodes and relationships are precise structured connections, which is the only plausible way the brain could store knowledge.
“The brain works not as a statistical engine but as a biological graph processor where concepts are nodes and relationships are precise structured connections”
The neocortex is composed of millions of nearly uniform repeating structures called cortical columns that all start with the same circuitry, and they specialize (e.g., visual cortex tuning to edges/colors, prefrontal cortex to planning) based on the signals they receive, implying the cortex is built from a single repeated general-purpose module.
“The Neoortex consists of millions of repeating structures called cortical columns. These aren't identical, but they're close.”
Modern artificial neural networks and predictive coding models assume smooth gradients, statistical functions, and uniform layers, and while good at pattern recognition, they fall short at reasoning, understanding exceptions, or explaining concepts—partly because biological neurons are too slow to operate that way.
“Modern artificial neural networks or even predictive coding models assume smooth gradients, statistical functions, and uniform layers. They can do impressive things like pattern recognition, but they fall short when you ask them to reason, understand exceptions, or explain concepts.”
Each cortical column functions like a node in a graph, with connections representing relationships (is-a, has-a, part-of) that support inheritance, exceptions, and bidirectional links, giving the brain meaningful symbolic reasoning that today's artificial neural networks cannot do.
“Each cortical column functions like a node in a graph.”
Each cortical column functions as a node in a graph, where the column itself represents a concept and connections to other columns represent relationships between concepts.
“I believe there's a compelling answer. Each cortical column functions like a node in a graph.”
The brain works not as a statistical engine but as a biological graph processor where concepts are nodes and relationships are precise structured connections, enabling representation of millions of concepts that are linked quickly in flexible and powerful ways.
“The brain works not as a statistical engine but as a biological graph processor where concepts are nodes and relationships are precise structured connections. And because the neoortex is full of these pre-allocated columns, we can represent millions of concepts and link them quickly in incredibly flexible, powerful ways.”
A cortical column needs only about 100 neurons because that is enough for input/output neurons, a dozen or so AND-gates, and supporting functions, which implies most of the brain's neurons are not there for brute-force learning but are scaffolding waiting to be given meaning through experience.
“It explains why columns need only about a 100 neurons. That's enough for input output a dozen or so andgates and the other functions I mentioned.”
Learning a relationship in the brain occurs by selectively strengthening a tiny subset of pre-existing near-zero-weight synapses (e.g., from the 'is-a' column to an AND-gate in the Fido column and from that gate to the 'dog' output neuron), so the physical structure is already in place and learning just activates it.
“Meaning is created by selectively strengthening a tiny subset of synapses. The physical structure is already in place and learning just activates it.”
In a graph-based knowledge model, each cortical column is a node like 'pho' or 'dog' with connections representing relationships like 'is a', 'has a', 'part of', etc., where nodes can inherit attributes, handle exceptions, and form bidirectional links.
“In this model, each column is a node like phto or dog with connections that represent relationships like is a has a part of and so on. These nodes can inherit attributes, handle exceptions and even form birectional links.”
Understanding how knowledge is structured in the brain reveals a blueprint for how AI could be dramatically more efficient, intelligent, and humanlike, addressing the fact that the human brain runs on just 20 watts while showing understanding and common sense far beyond today's AI systems.
“When we really understand how knowledge is structured in the brain, we're uncovering a blueprint for how AI could be dramatically more efficient, intelligent, and humanlike. Your brain runs on just 20 watts, yet it shows understanding and common sense far beyond today's AI systems.”
Cortical columns are not just interesting anatomy—they are the nodes in the graph of thought, the fundamental units implementing conceptual understanding and reasoning.
“Cortical columns aren't just interesting anatomy. They are the nodes in the graph of thought.”
Cortical columns are not merely anatomical structures of interest but are fundamental to understanding how the brain produces thought, serving as the nodes implementing the graph of conceptual knowledge and relationship processing.
“Cortical columns aren't just interesting anatomy. They are the nodes in the graph of thought.”
Modern artificial neural networks and predictive coding models assume smooth gradients, statistical functions, and uniform layers, which enables impressive pattern recognition but causes them to fall short when reasoning, understanding exceptions, or explaining concepts.
“Now, here's where this totally breaks from traditional AI. Modern artificial neural networks or even predictive coding models assume smooth gradients, statistical functions, and uniform layers. They can do impressive things like pattern recognition, but they fall short when you ask them to reason, understand exceptions, or explain concepts.”
The graph-based model explains why cortical columns need only approximately 100 neurons: enough for input/output neurons, roughly a dozen AND gates, and the other supporting functions, suggesting most brain neurons are scaffolding waiting to be given meaning through experience rather than involved in brute-force learning.
“This is such a powerful model. It explains why columns need only about a 100 neurons. That's enough for input output a dozen or so andgates and the other functions I mentioned. It also suggests that most of the brain's neurons are not there for brute force learning. They're the scaffolding waiting to be given meaning through experience.”
Knowledge in the graph model is created by selectively strengthening synapses based on pre-existing near-zero-weight connections, not by creating entirely new connections, which means the brain's structural capacity is essentially fixed and learning is about activating latent structure.
“Meaning is created by selectively strengthening a tiny subset of synapses. The physical structure is already in place and learning just activates it.”
Most synaptic connections inside cortical columns and between neighboring columns already exist at near-zero weight, meaning knowledge is created by selectively strengthening a tiny subset of synapses, and the physical structure is already in place—learning simply activates it.
“most of the synaptic connections inside the column already exist along with the connections to neighboring columns. They just start with a near zero weight. Meaning is created by selectively strengthening a tiny subset of synapses. The physical structure is already in place and learning just activates it.”
The graph structure enables meaningful symbolic reasoning where activating a concept node like 'pho' retrieves all inherited attributes of the parent concept 'dog' unless exceptions override them, and vice versa—activating 'dog' can discover its instances like 'pho'—capabilities that artificial neural networks cannot provide.
“For example, if phto was a dog, we can activate phto and retrieve all the attributes of dogs unless there's an exception. And we can also activate dog and discover its instances like pho. This graph structure gives us meaningful symbolic reasoning, something today's artificial neural networks just can't do.”
To build AI that actually thinks and reasons like humans, researchers need to shift from layering up weights in statistical models to wiring in explicit structure, implementing graph-based symbolic processing instead of gradient-based learning.
“If we want to build AI that actually thinks, we need to stop layering up weights and start wiring in structure.”
Biological neurons are too slow to perform gradient-based learning like artificial neural networks, making cortical columns incompatible with statistical learning models and instead fitting perfectly with graph-based knowledge models.
“Cortical columns don't fit that mold, partly because biological neurons are just too slow to be used that way. Instead, cortical columns fit perfectly with a graphbased knowledge model.”
The neocortex can be reorganized into conceptual layers with output neurons near the top that fire when a concept is the result of a query, and input neurons in lower layers where query inputs occur, enabling the activation patterns described in the previous video.
“Let's take our previous example of phto is a dog and a dog has fur and rearrange the neurons so they form into the conceptual layers of the neoortex. Near the top of each column is the output neuron. In this case, it's the phto output or the dog output. These are the neurons which fire in the event that phto or dog is the result of a query. In a layer below are the input neurons where query inputs happen.”
Learning a relationship like 'pho is a dog' requires strengthening two key synapses: one from the 'isa' column to an AND gate in the 'pho' column and another from that gate to the output neuron in the 'dog' column, with mirror connections strengthened to enable the inverse relationship 'dog is instance of pho'.
“When learning this relationship, we need to strengthen two key synapses. one from the is a column to an and gate in the phyto column and another from that gate to the out neuron in the dog column. To make the inverse relationship work, we also strengthen a mirror connection from dog to phto with the relationship type set to is instance of”
Additional neurons within cortical columns support specialized functions including: firing the spike bursts needed for Hebbian learning, setting up inverse relationships, ensuring columns are allocated only once until meaning is forgotten, handling exceptions, and managing local redundancy for robustness and noise resistance.
“Additional neurons within the column support functions like firing the bursts needed for heavy in learning, setting up inverse relationships, ensuring that a column is only allocated once and not reused until its meaning is forgotten, handling exceptions, and managing local redundancy for robustness and noise resistance.”
If understanding the brain's knowledge structure is correct, it explains how the brain achieves understanding, generalization, and common sense, while explaining why current AI lacks these abilities because it does not implement graph-based symbolic processing.
“So here's the bottom line. If you're wondering how the brain achieves understanding, generalization, and common sense, this is how. If you're wondering why today's AI lacks those abilities, it's because it doesn't do this.”
If your brain needs to know what 'pho' is, it fires the 'pho' input neuron and the 'is a' neuron, and the 'dog' output neuron will fire as a result of the strengthened synaptic connections representing this relationship.
“if your brain needs to know what pho is, it fires the pho input neuron and the is a neuron and the dog output neuron will fire.”
To store a fact like 'pho is a dog', the brain must allocate a new cortical column or locate an already defined column for pho, dog, and 'isa' relationship, each containing an input neuron, output neuron, pre-wired neurons acting as gates representing relationships, and supporting neurons that can fire the spike bursts needed for Hebbian learning.
“Let's say we want to store the fact that phto is a dog. We need to allocate a new or locate an already defined column for pho dog and isa. These columns each include an in neuron and an out neuron plus a pool of pre-wired neurons acting as gates used to represent relationships and the supporting neurons which can fire the necessary spike bursts to implement the heavy in learning.”
This graph-based knowledge structure provides incredible information compression, which is something both the brain and artificial intelligence systems need.
“This structure enables incredible compression of information just like the brain needs and AI does, too.”
In the previous video, Simon showed how knowledge can be represented as a graph of relationships including inheritance and exceptions, where a concept like 'pho is a dog' can be represented with just a few neurons, and attributes like 'has fur' are inherited unless overridden by exceptions.
“In the last video, I showed how knowledge can be represented as a graph of relationships complete with inheritance and exceptions. We can represent the concept that pho is a dog with just a few neurons. Further, if a dog has fur, then phto inherits the has fur attribute on the fly. But if stubby is a dog and dogs have tails, yet Stubby has no tail, the exception overrides just that attribute.”
The brain has three major parts: the brain stem managing vital functions like heartbeat and breathing, the cerebellum handling coordination and motor learning, and the neocortex where conscious thought, planning, and reasoning happen.
“The brain stem, which keeps us alive by managing things like heartbeat and breathing. the cerebellum which handles coordination and motor learning and at the top the neoortex where conscious thought planning and reasoning happen.”
The brain has three major parts: the brainstem managing autonomic functions like heartbeat and breathing, the cerebellum handling coordination and motor learning, and the neocortex where conscious thought, planning, and reasoning occur.
“The brain has three major parts. The brain stem, which keeps us alive by managing things like heartbeat and breathing. the cerebellum which handles coordination and motor learning and at the top the neoortex where conscious thought planning and reasoning happen.”
Cortical columns are small structures containing approximately 100 neurons each, vertically organized through all six layers of the cortex, with each column functioning as a tiny processing unit.
“These columns are small about a 100 neurons each and vertically organized through all six layers of the cortex. Each one acts like a tiny processing unit.”
Connections between nearby cortical columns can use shorter lateral connections, while connections between distant columns use longer connections that dive into white matter deeper in the brain, a distinction which computer simulations don't need to make as all connections are treated as lateral.
“If a connection is needed between nearby columns, the brain can utilize one of the shorter lateral connections. If the connection is between columns at a greater distance, the connection uses one of the longer connections which dives into the white matter deeper in the brain. In our computer simulations, connections are represented by a few numbers in an array. So we don't need to make this distinction and we treat all connections as lateral.”
The neocortex, if unwrapped, would form a ragged disc approximately one foot across, consisting of gray matter packed with neurons arranged in six layers on one surface and white matter containing axons below.
“Let's talk about the neoortex, the folded outer surface of your brain. If you unwrapped it, it would form a ragged disc about a foot across. On one surface would be gray matter packed with neurons arranged in six layers. And below that is white matter, a tangled mass of axons connecting everything together.”
Bidirectional links allow the model to activate Fido and retrieve all dog attributes (unless overridden by exceptions), and also to activate 'dog' and discover its instances like Fido, by strengthening a mirror connection with the relationship type 'is instance of'.
“if phto was a dog, we can activate phto and retrieve all the attributes of dogs unless there's an exception. And we can also activate dog and discover its instances like pho.”
The white matter in the brain resembles an old plugboard from the early days of computing, suggesting an analogy between biological neural architecture and early electronic computing logic.
“It resembles an old plugboard from the early days of computing.”
Simon uses open-source brain simulator projects for demonstrations and simulations throughout this video series, which viewers can download from GitHub and use to experiment with neural circuits.
“I'm using our open-source brain simulator projects for simulations and demonstrations throughout this video series. If you want to experiment with your own neural circuits, you can download the brain simulator 2 from GitHub and try out these neural circuits.”
Charles Simon is a longtime AI researcher and software developer with experience developing software for neurological test instruments and neural simulators.
“I'm Charles Simon, longtime AI researcher, software developer, and manager. Beyond AI, I've developed software for neurological test instruments and neural simulators.”
Charles Simon created the Future AI Society to explore how neuroscience can inform smarter, more humanlike AI development.
“I created the Future AI Society to explore how neuroscience can inform smarter, more humanlike AI.”
In computer simulations of this model, connections are represented as a few numbers in an array, so the lateral-versus-white-matter distinction between short and long connections need not be made and all connections are treated as lateral.
“In our computer simulations, connections are represented by a few numbers in an array. So we don't need to make this distinction and we treat all connections as lateral.”