I'm Kohen a User

Latent Spaces and the Brain’s Dynamic Geometry

Category:
AI & Technology
Published:

This image reflects the tone and underlying structure of the article.

How Waves Implement the “Platonic Forms” on Wetware

Note to the Reader
This essay explores how “Meaning” and “Consciousness” might be established as spatial structures, using the Latent Space of Large Language Models and Spatial Computing in neuroscience as our guides.
Please note that “Spatial Computing” here refers not to XR (VR/AR) technology, but to the neuroscientific concept where the brain utilizes physical location on the cortex as a computational resource.

How the Brain Implements Platonic Structures

When we visualize the interior of a Large Language Model, we are often met with a contradiction: a sense of dissonance mixed with overriding awe.
“Meaning” itself is nowhere to be found. Instead, we see only a constellation of points suspended in a high-dimensional vector space (often tens of thousands of dimensions) defined entirely by their distance relationships. Concepts do not appear to be stored as discrete data items; rather, they manifest as distributions.

What is truly strange is that even with different model architectures or training data, once performance reaches a certain threshold, the structure of these internal representations becomes surprisingly similar. Whether the model learns languages, images, molecules, or physical simulations, the topography of the resulting high-dimensional space tends to converge toward a remarkably consistent shape.

A hypothesis proposed to explain this is “The Platonic Representation Hypothesis.” It suggests that meaning and function may not be arbitrary symbols. Rather, they can be seen as real, ideal geometric structures (Platonic forms) that any intelligence must inevitably approximate in order to understand the world.

Spatial Computing in the Brain

This brings us to a biological conundrum.

AI can hold these Platonic structures directly as high-dimensional mathematical spaces (matrices in memory). However, our brain is a 3-dimensional object bound by physical constraints. It is metabolically active, noisy, and possesses a limited physical surface area. How, then, does the cortex implement the kind of high-dimensional geometry that AI possesses?

To address this, researchers at MIT, led by Professor Earl Miller, have proposed a framework called “Spatial Computing.”
They challenge the view that the brain computes solely through fixed neuronal wiring. Instead, they propose that the brain uses neural oscillations (brainwaves) to dynamically control physical space, instantiating the stage for computation on the fly.

Converting Space and Time

To understand how this might work, we must look at how the brain uses “time” to navigate the constraints of “space.”
It is impossible to etch a complex, high-dimensional structure onto the physical surface of the brain all at once. Therefore, the brain utilizes rhythms such as alpha and beta waves.

We might think of this as a stencil: a wave of a specific frequency temporarily isolates a region on the cortex, creating a pathway for information. Within that open window, a temporary computational circuit emerges via gamma waves.
In this model, the brain does not unfold the high-dimensional structure simultaneously. Instead, it appears to use the time axis of neural oscillations to project different “cross-sections” of the structure onto physical space, moment by moment.

Static Parameters and Dynamic Waves

Here, a theoretical connection emerges between the latent space of Artificial Intelligence and the Spatial Computing of the brain.

The Platonic Representation Hypothesis shows us what should be represented: the Goal (Mathematical Structure) of how meaning should be arranged. Conversely, MIT’s Spatial Computing offers a perspective on how a physical brain implements it: the Means (Physical Implementation).

  • Meaning in AI exists as a structure frozen in the static medium of learned parameters.
  • Meaning in the Brain, by contrast, appears to arise as a dynamic wave, temporarily drawn onto physical space by the flow of energy.

What both share is the destination: the Topology. They differ primarily in how they preserve it.

From this perspective, the brain’s ability to bridge physical distances does not require a quantum explanation. It is better understood as “Dynamic Functional Connectivity” mediated by oscillations.

For a coherent structure to emerge, neurons do not need to be physically wired to every other neuron. When distant regions lock their phases through the synchronization of brainwaves, they act as a single functional unit despite the spatial separation.
(Note: Unlike Integrated Information Theory, which quantifies the causal irreducibility of such states, our focus here is strictly on the mechanism of how this spatial alignment is dynamically constructed.)

In this view, the brain overcomes physical distance not through “spooky action,” but through temporal coherence. What binds the information is not a physical cable, but a shared rhythm that sustains the topology across the cortex.

Consciousness: Not a “Viewer,” But a “State”

When we talk about consciousness, we often rely on an intuitive model: “Information is integrated somewhere, and someone is watching it.”
However, this naive model implicitly assumes a place where integration happens or a subject who watches: a dedicated “integration node” or an internal observer (a homunculus). Modern science, of course, has found no such tiny center.

If we define meaning itself as a spatial structure (or its dynamic projection), however, the picture changes. We may be able to explain the state we call “conscious” without positing a dedicated module.

Instead of attributing consciousness to a central hub or an observer, consciousness may be better understood as a physical state where a semantic geometric structure is established with a certain degree of stability and complexity.
In this view, consciousness is not the “viewer”; it is the “state being drawn” by the waves in this very moment.

In fact, AI research uses methods like CKA (Centered Kernel Alignment) and RSA (Representational Similarity Analysis) to quantitatively evaluate the structural similarity between different models. The connection point between the Brain and AI may not be in mimicking the neuron as a component, but in aligning this “topology of the activation space.”

Epilogue

Meaning is not something preserved in matter. It is something drawn in space, each time anew, as a wave.

To transcribe the universal structures ubiquitous in the universe (Platonic forms) onto physical space for just a fleeting moment.
That ephemeral drawing process may be the true nature of the phenomenon we call “intelligence,” and what we feel as “I.”

Related Resources

This article is an English translation of the original Japanese post.
Title: AIの潜在空間と脳の空間計算
Published: February 2, 2026
Original Source: Zenn.dev

Post Navigation

Search This Site