Get AI summaries of any video or article — Sign up free
Your Brain Is 150,000 Mini-Brains thumbnail

Your Brain Is 150,000 Mini-Brains

Artem Kirsanov·
6 min read

Based on Artem Kirsanov's video on YouTube. If you like this content, support the original creators by watching, liking and subscribing to their content.

TL;DR

Neocortex columns show both shared six-layer anatomy and consistent functional tuning, supporting the idea of repeated processing modules.

Briefing

The neocortex may work less like a patchwork of specialized brain regions and more like a single repeating computation: thousands of “mini-brains” (cortical columns) each learn a predictive model by binding sensory inputs to location signals generated through movement. The payoff is a unified explanation for how the same neural machinery can support perception, language, and abstract reasoning—without requiring evolution to invent and wire together hundreds of different algorithms.

For more than a century, researchers have found that the neocortex’s anatomy follows a common blueprint: the same six-layer circuitry and cell types recur across the cortical sheet. Functionally, experiments strengthened the case for a “columnar” principle. When electrodes were lowered through the cortical layers, neurons within a vertical column—roughly half a millimeter wide—responded to the same patch of skin in the somatosensory cortex. A similar pattern emerged in the visual system, where columns were tuned to lines with particular orientations. Taken together, the evidence suggests the cortex is built from repeated processing modules rather than many unrelated structures.

The thousand brains theory (pioneered by Jeff Hawings and the Nenta team) turns that anatomical regularity into a computational claim. If each column shares the same functional hardware, then each column likely runs the same universal algorithm. Instead of the brain constructing one centralized model of the world, each cortical column builds its own complete predictive model. Conscious experience then emerges from rapid cross-column “voting,” where competing hypotheses reinforce one another when they agree and get suppressed when they don’t.

A key mechanism for building a model is sensory–motor coupling. The theory uses a dark-room analogy: touching an unseen statue with only fingertip sensations yields disconnected features (smooth, sharp, rough) but no coherent object. When the brain also tracks movement—updating a mental reference frame via path integration—those features become anchored to locations. With that reference frame, a column can predict what sensations should occur if the sensor moves again.

The theory further proposes that this sensory–motor loop is implemented across the neocortex’s six layers. Layer 4 acts as the primary input port for “what” information, receiving sensory data. Layer 6 is predicted to contain grid cell–like neurons that track “where” by performing path integration using movement-related signals. Layer 5 generates motor outputs and also sends a copy of motor commands upward so layer 6 can update its internal coordinates. Layers 2–3 then bind sensory features with location context to form object-level representations, while layer 1 helps prime dendritic activity to set up predictions.

To explain why this architecture might have appeared so quickly in evolution, the theory offers a simple scaling story: evolution didn’t need to invent many new cognitive components; it could expand intelligence by adding more instances of the same columnar module. Finally, the same algorithm is claimed to scale upward into abstract thought. Higher-level columns can treat the outputs of lower columns as their “sensory input,” and their “motor commands” can correspond to shifts in attention or memory retrieval—turning navigation through physical space into navigation through conceptual space.

Cornell Notes

The thousand brains theory treats the neocortex as a set of repeating “mini-brains” (cortical columns) that all run the same universal algorithm. Each column builds a predictive model by binding sensory features (“what”) to a location-based reference frame (“where”) computed through sensory–motor coupling and path integration. The theory maps this loop onto the neocortex’s six layers: layer 4 supplies sensory input, layer 6 provides grid-like coordinate tracking, layer 5 generates motor output (and sends a copy for updating coordinates), and layers 2–3 bind features with context to form stable object representations. Perception becomes unified through voting: columns exchange sideways signals so mutually consistent hypotheses reinforce and inconsistent ones fade. The same mechanism is proposed to extend from physical perception to abstract reasoning by swapping physical sensors/movements for attention and memory shifts.

What evidence supports the idea that the neocortex is organized into repeating columns rather than many unrelated circuits?

Anatomical studies found a shared blueprint across the cortical sheet: the same six-layer arrangement and cell types recur throughout the neocortex. Functionally, microelectrode recordings showed that neurons within a vertical column (about half a millimeter wide) respond to the same patch of sensory input in the somatosensory cortex as electrodes pass through layers. A parallel result appeared in the visual cortex, where columns were tuned to specific line orientations, reinforcing the “columnar principle.”

How does the thousand brains theory connect predictive modeling to cortical columns?

Modern predictive accounts emphasize that brains build internal models to anticipate what comes next. The thousand brains twist is that there isn’t just one global model: every cortical column is its own modeling system. Each column uses sensory–motor coupling to learn structure, then predicts future sensory consequences of movement. The final percept arises when columns compare notes—through voting—so only hypotheses consistent across columns become dominant.

Why is movement essential in this framework, and what is “path integration” doing?

Sensation alone can list features without tying them to a coherent object. Movement supplies the missing structure by updating a reference frame. In the theory’s dark-room statue analogy, tracking hand motion lets the brain bind tactile features to locations derived from movement. This location tracking is described as path integration, the computation that supports navigation in the dark and returning to where you started.

How does the theory map “what” and “where” signals onto specific cortical layers?

Layer 4 is treated as the “what” input port, receiving sensory data first when edges or textures are detected. Layer 6 is predicted to generate “where” via grid cell–like neurons that track coordinates using movement information. Layer 5 provides motor output and also sends an internal copy of motor commands to layer 6 so coordinates can be updated. Layers 2–3 then bind sensory features with location context to represent objects, while layer 1 helps prime dendritic predictions using positional information.

What does “voting” mean in practice, and how does it resolve ambiguity?

A single sensory input can support multiple interpretations. When touching a coffee cup in the dark, different fingers activate different sets of columns representing competing objects (e.g., coffee cup vs. saucer vs. wine glass). The theory says columns exchange sideways signals so hypotheses that match across columns reinforce each other. The coffee cup hypothesis is the overlap: it receives supporting evidence from both the handle-touching thumb columns and the rim-touching index columns, so it strengthens while inconsistent hypotheses weaken and get suppressed.

How does the same algorithm scale from physical perception to abstract reasoning?

The theory claims the algorithm stays the same while the “sensor” and “motor” change. For higher-level cognition, layer 4 input need not come from skin or eyes; it can come from the outputs of lower-level columns. Likewise, motor commands can correspond to shifting attention or recalling memories. In that view, solving a logic problem becomes navigating an abstract space, with each mental step moving to a new “location” in conceptual terrain.

Review Questions

  1. What specific role does path integration play in turning raw sensory features into a structured model?
  2. How do layers 4, 5, 6, and 2–3 each contribute to the sensory–motor loop described by the thousand brains theory?
  3. Why does the theory predict that adding more cortical columns can increase intelligence without inventing new cognitive modules?

Key Points

  1. 1

    Neocortex columns show both shared six-layer anatomy and consistent functional tuning, supporting the idea of repeated processing modules.

  2. 2

    The thousand brains theory proposes that each cortical column runs the same universal predictive algorithm rather than the cortex acting as one monolithic model-builder.

  3. 3

    Sensory–motor coupling is central: movement updates a reference frame so sensory features can be bound to locations.

  4. 4

    Layer 4 is treated as the main “what” input channel, while layer 6 is predicted to provide grid-like “where” tracking using movement-derived signals.

  5. 5

    Layer 5 generates motor outputs and also provides an internal copy of motor commands that helps layer 6 perform path integration.

  6. 6

    Layers 2–3 bind sensory features with location context to form stable object-level representations, with layer 1 priming predictions.

  7. 7

    Unified perception emerges through voting: hypotheses consistent across columns reinforce each other while inconsistent interpretations fade quickly.

Highlights

A column can’t build a coherent model from sensations alone; binding features to a movement-updated reference frame is what turns “lists of features” into structured understanding.
The theory assigns a concrete division of labor across layers: layer 4 for sensory “what,” layer 6 for coordinate “where,” layer 5 for motor output (plus a copy), and layers 2–3 for binding into object representations.
Voting across columns resolves ambiguity fast—only hypotheses that overlap across multiple finger inputs (or other sensor streams) survive.

Topics

  • Cortical Columns
  • Predictive Modeling
  • Path Integration
  • Grid Cells
  • Neocortex Voting

Mentioned

  • Jeff Hawings
  • Rammonica Hall
  • Vernon MCastle
  • David Hubil
  • Torstston Weasel