Intrinsic Dimension of Neural Representations

This direction is merged into another direction: intrinsic dimension estimation. Its last version stays readable and citable here.

Version 1 of 121 papers21 conceptsfrozen 2026-10-04

When we look at a modern Large Language Model, we see vectors with thousands of dimensions. A single token might be represented by a 4,096-element array. However, researchers have increasingly realized that these high-dimensional spaces are mostly empty. The actual "meaning" or information within these layers lives on a much smaller, curved surface called a manifold.

The Intrinsic Dimension (ID) is a measure of the minimum number of variables required to describe this manifold without losing information. By tracking how ID changes across the layers of a network, we can see the "geometric signature" of how a model thinks, generalizes, and abstracts.

The Geometry of the Hidden Lifecycle

In many deep learning architectures, information follows a predictable geometric trajectory. Abhinav Joshi and colleagues describe this as a "hunchback" pattern: as data flows through the model, the intrinsic dimensionality first expands to process the complexity of the input and then sharply compresses as the model "commits" to a specific decision or output.

This lifecycle is particularly visible in protein language models (pLMs). Lucrezia Valeriani and her team identified a "three-phase" strategy:

  1. Expansion: The model projects raw input into a higher-dimensional feature space.
  2. Distillation: A low-dimensional plateau in the middle layers where the most critical biological information is concentrated.
  3. Reconstruction: A final ascent in dimensionality to map these abstract features back to the vocabulary of the output.

This suggests that the "sweet spot" for extracting useful features isn't at the very end of a model, but in the middle, where the intrinsic dimension is often at its most refined.

Measuring Complexity Without Labels

How do we calculate the dimension of a manifold we can't see? The most common intuition relies on local distances. If you double the radius of a circle (1D), its "volume" (length) doubles. If you double the radius of a sphere (2D), its volume (area) quadruples.

By measuring the ratio of distances between a data point and its nearest neighbors—a technique used by Pascal Tikeng Notsawo and others—we can estimate the effective degrees of freedom. This allows researchers like Arjun Rao to evaluate "geographic embeddings" (vectors representing locations on Earth). Even if a vector is 512 dimensions wide, its ID might be as low as 2 or 10, revealing exactly how much information density the model has actually captured about the physical world.

Why Dimension Matters for Generalization

Intrinsic dimension isn't just a descriptive statistic; it is a predictor of model behavior.

Predicting Grokking

One of the most mysterious phenomena in AI is "grokking," where a model suddenly learns to generalize after a long period of overfitting. Pascal Tikeng Notsawo found that the "spectral signature" of the activation manifold's dimensionality can predict whether a model will eventually "grok" long before it actually happens. Essentially, the geometry of the internal representations begins to simplify and organize itself well before the loss curve shows any improvement.

Brain Alignment and Abstraction

Why do the middle layers of AI models often predict human brain activity better than the final layers? Emily Cheng argues that this "brain-model alignment" is driven by meaning abstraction. As models learn to build rich, abstract representations, their intrinsic dimension peaks in the intermediate layers. This peak in complexity represents the model's attempt to integrate various features into a coherent concept—a process that mirrors how the human cortex organizes information.

Robustness and Efficiency

Understanding where the manifold is most compact can also make models more efficient. Enes Altinisik and Safa Messaoud leveraged ID to improve adversarial training. By identifying the specific layers where the representation is most compressed (the lowest ID), they developed SMAAT, a method that targets those layers for intervention. This approach boosts the model's robustness against attacks while significantly cutting training time.

A Universal Lens

Whether applied to the human gut microbiome—where Ivan Bulygin used ID to show that microbial communities exist on a continuous spectrum rather than discrete "types"—or to the role of punctuation in LLMs, intrinsic dimension provides a task-agnostic way to measure information.

As Saahith Janapati demonstrated, even the way we prompt a model changes its geometry: In-Context Learning (ICL) induces higher-dimensional, more complex representations than standard fine-tuning. By looking at the world through the lens of intrinsic dimension, we move past looking at what a model predicts and begin to understand the shape of its knowledge.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_intrinsic_dimension_of_neural_representa_9ac084d1,
  title = {Intrinsic Dimension of Neural Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intrinsic-dimension-of-neural-representations/rdr_9ac084d12df745cc8f52e01248c39c5a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_intrinsic_dimension_of_neural_representa_9ac084d1_v1,
  title = {Intrinsic Dimension of Neural Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intrinsic-dimension-of-neural-representations/rdr_9ac084d12df745cc8f52e01248c39c5a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}