The Expansion and Compression of Neural Representation Geometry

This direction is split into other directions: intrinsic dimensionality peak. Its last version stays readable and citable here.

Version 1 of 126 papers35 conceptsfrozen 2026-10-04

Deep neural networks are often viewed as "black boxes" that transform inputs into outputs through a series of opaque linear and non-linear operations. However, recent research into the intrinsic dimensionality (ID) of hidden representations—the minimum number of variables needed to describe the data manifold—reveals a surprisingly consistent geometric logic.

Across diverse architectures like Transformers, RNNs, and CNNs, information does not simply flow linearly. Instead, it follows a "hunchback" trajectory: expanding into high-dimensional spaces to disentangle features before aggressively compressing into a low-dimensional manifold to make a final decision.

The Hunchback Profile

When researchers measure the ID of data as it passes through a network, they consistently observe a non-monotonic curve. Akshay Rangamani and Tomaso Poggio describe this as a "hunchback" dimensionality profile.

In the early layers, the dimensionality of the representation increases. In the middle layers, it reaches a peak. In the final layers, it contracts sharply. This pattern suggests that deep learning is a two-part process: first, the model "unfolds" the data to understand its complexity, and then it "distills" that complexity into a usable answer.

Phase 1: Expansion and Disentanglement

Why would a network increase the dimensionality of its internal representations? The intuition is that raw data (like pixels or word embeddings) is often "tangled." For a network to distinguish between classes, it must project this data into a higher-dimensional space where different features can be linearly separated.

Matthew Farrell and Eric Shea-Brown argue that in recurrent networks, chaotic internal dynamics act as an expansionary force, allowing the network to explore a wide range of structural variations Dynamic compression and expansion in a classifying recurrent network. This expansion is crucial for "pre-processing" the data, ensuring that the model captures as many relevant features as possible before attempting to generalize.

Phase 2: The High-Dimensional Abstraction Peak

The point of maximum intrinsic dimensionality is not just a mathematical curiosity; it represents a functional pivot. Emily Cheng, Diego Doimo, and Alessandro Laio have identified this as a "high-dimensional abstraction phase".

At this peak, the model begins to discard "surface-level" features—literal details like sentence length or specific word frequencies—in favor of abstract concepts like syntax and semantics. Abhinav Joshi and colleagues found that immediately after this peak, the model's internal representations align much more closely with the final correct output, suggesting the peak is the moment the model "commits" to a conceptual understanding Geometry of Decision Making in Language Models.

Phase 3: Compression and Neural Collapse

Once the model has abstracted the necessary information, it enters a compression phase. The goal here is to map the high-dimensional concepts onto a low-dimensional output space (like a classification label or the next token in a sequence).

This compression is often linked to the phenomenon of Neural Collapse. Robert Wu and Vardan Papyan show that as models scale, their internal representations "collapse" into clean, geometric structures centered around class means Linguistic Collapse: Neural Collapse in (Large) Language Models. By reducing the number of dimensions required to represent the data, the model becomes more robust and generalizes better to unseen examples.

Geometric Evolution During Training

This expansion-compression cycle doesn't just happen across layers; it also happens across training time. Melody Zixuan Li and Guillaume Lajoie describe a "triphasic" trajectory during pre-training:

  1. Initial Collapse: The model suppresses random initialization noise.
  2. Expansion: The model begins to memorize data statistics and pack complex concepts into its activation space.
  3. Final Compression: The model distills these concepts into structured features for generalization Tracing the Representation Geometry of Language Models from Pretraining to Post-training.

Tools like TRACE, developed by Nura Aljaafari and Andre Freitas, allow researchers to monitor these phase transitions in real-time. By watching the intrinsic dimensionality stabilize, they can pinpoint the exact moment a model shifts from rote memorization to true linguistic understanding TRACE: Training and Inference-Time Interpretability Analysis for Language Models.

Why It Works

The success of this geometric strategy lies in the trade-off between expressivity and robustness. High dimensionality provides the expressivity needed to disentangle complex, overlapping features in the input. Low dimensionality provides the robustness needed to ignore noise and focus on the underlying semantic manifold. By navigating this expansion-compression trajectory, neural networks effectively act as internal autoencoders, finding the "sweet spot" of semantic information in their middle layers.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_the_expansion_and_compression_of_neural__aa7e0ba8,
  title = {The Expansion and Compression of Neural Representation Geometry},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-expansion-and-compression-of-neural-representation-geometry/rdr_aa7e0ba82b344aafb4391524a8faba75},
  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_the_expansion_and_compression_of_neural__aa7e0ba8_v1,
  title = {The Expansion and Compression of Neural Representation Geometry},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-expansion-and-compression-of-neural-representation-geometry/rdr_aa7e0ba82b344aafb4391524a8faba75/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}