🤖 machine learning

Transformer Geometry Observatory TGO-IV: Developmental Topology Observatory

This paper introduces Transformer Geometry Observatory (TGO-IV), a novel framework that utilizes Persistent Homology and Vietoris–Rips simplicial complexes to comprehensively analyze the topological evolution of Transformer representations across layers, thereby offering deeper insights into how raw inputs transform into task-relevant features.

Original authors: Kaustubh Kapil, Kishor P. Upla

Published 2026-08-12

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Kaustubh Kapil, Kishor P. Upla

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Shape of Thinking: How AI Builds Its World

Imagine you are trying to understand how a giant, digital brain learns to recognize a cat. For a long time, scientists treated these "Transformers"—the powerful AI models behind your favorite chatbots and image generators—as mysterious black boxes. We knew they worked incredibly well, but we didn't really know how they learned. To peek inside, researchers started looking at the "representations" the AI creates. Think of these representations as a cloud of glowing dots floating in a high-dimensional space. Each dot is a piece of the image (like a patch of fur or an ear), and the way the dots are arranged tells the AI what it's seeing.

Previously, scientists studied these clouds by measuring their "spectrum" (how spread out the dots are) or their "geometry" (how the dots cluster together). They found that as the AI learns, the cloud expands and organizes itself into meaningful shapes. But there was a missing piece of the puzzle: the topology. In simple terms, topology is the study of shapes that don't care about stretching or bending, but do care about holes and connections. Is the cloud one big solid blob? Is it a ring with a hole in the middle? Is it a bunch of disconnected islands? Understanding the "shape" of these clouds helps us see if the AI is truly understanding the structure of the world or just memorizing patterns. This is the question that the new study, TGO-IV, sets out to answer.

The Paper: TGO-IV, the Topology Detective

In this new study, titled TGO-IV: Developmental Topology Observatory, researchers Kaustubh Kapil and Kishor P. Upla from India's SVNIT take a magnifying glass to the "shape" of how a Vision Transformer (a type of AI that sees images) learns. They are part of a larger team called the Neuromorphic Intelligence Research Collective, and this is the fourth installment in their "Transformer Geometry Observatory" series. While the previous parts looked at the math of the AI's spread and its semantic meaning, TGO-IV asks a deeper question: Does the global shape of the AI's thinking change as it learns, or does it stay the same?

To answer this, the team didn't just look at the dots; they built a digital "scaffolding" around them. Imagine you have a cloud of 197 glowing dots (representing the parts of an image) floating in a 384-dimensional space. You can't see the shape just by looking at the dots. So, the researchers used a mathematical trick called a Vietoris–Rips simplicial complex. Think of this as a game of "connect the dots." They start by drawing a line between any two dots that are very close together. Then, they slowly increase the size of the "connection radius." As the radius grows, lines connect more dots, triangles form between three connected dots, and eventually, 3D shapes (like tetrahedrons) appear. This creates a mesh that approximates the shape of the data cloud.

By watching how this mesh changes as the radius grows, they can track "persistent homology." This is a fancy way of counting the life stories of holes and connections.

Birth: A new connection forms, or a loop appears.
Death: A connection gets so big that a hole gets filled in, or a loop closes up.

The researchers tracked these "births" and "deaths" as the AI processed an image through its 12 layers, and they did this over 100 days of training on 1,000 images from the ImageNet-100 dataset. They used a specific model, ViT-Small/16, which processes images by breaking them into 16x16 patches, resulting in 197 tokens (dots) per image.

What They Found: The Great Consolidation

The study revealed a fascinating story of how the AI's internal world stabilizes. Here is what the "topological observatories" saw:

1. The Islands Merge into a Continent
At the very beginning of training (Epoch 1), the AI's representation cloud was a scattered archipelago. The dots were far apart, forming many isolated islands. As training progressed, the researchers noticed a massive shift in the Betti curves (a graph that counts how many connected pieces exist). The "zeroth Betti number" ( $\beta_0$ ), which counts the number of separate islands, dropped dramatically. The dots started merging into fewer, larger groups much earlier in the process. By the end of training (Epoch 100), the cloud had become a single, cohesive continent. The AI wasn't just seeing parts; it was seeing the whole picture as one connected unit.

2. The Rings Stay, The Holes Fade
The researchers also looked for "loops" (rings of dots) and "voids" (3D bubbles). They found that while the AI created some loops ( $\beta_1$ ), they were always few and far between. Even more interestingly, the 3D bubbles ( $\beta_2$ ) were almost entirely absent. This suggests that the AI's thinking doesn't form complex, multi-layered bubbles of information. Instead, it focuses on connecting the dots into a solid, flat(ish) structure.

3. The Layers Get More Alike
One of the most surprising findings was about the distance between layers. The researchers used two tools, Bottleneck Distance and Wasserstein Distance, to measure how different the shape of the data was from one layer to the next.

Early Training: The layers were very different from each other. The shape of the data changed wildly as it moved from the first layer to the second, then to the third.
Late Training: As the AI learned, the layers started to look more and more similar. The "distance" between the shapes of consecutive layers shrank. It's as if the AI stopped reinventing the wheel at every step and started refining the same stable structure.

However, they did notice a few "peaks" where the shape changed significantly, specifically around Layers 3, 10, and 12. Even after the AI had "learned," these specific layers continued to do heavy lifting, reorganizing the data in significant ways.

The Big Idea: Progressive Topological Stabilization

Based on these observations, the authors propose a new idea called the Progressive Topological Stabilization Hypothesis.

They suggest that learning isn't about the AI's internal world collapsing into a boring, featureless blob. Instead, it's a process of refinement.

The Metaphor: Imagine a chaotic crowd of people at a concert. At first, everyone is scattered, shouting, and disconnected (high $\beta_0$ , many islands). As the music starts and the song builds, people start linking arms, forming small groups, and eventually, the whole crowd moves as one giant, cohesive wave. The "shape" of the crowd becomes stable and connected, but it doesn't disappear.
The Finding: The AI's representations become "increasingly compact" (the dots get closer) while keeping their "non-trivial geometric organization" (the structure remains interesting and useful). The AI doesn't just memorize; it builds a stable, connected map of the world.

What This Means (and What It Doesn't)

The authors are careful to state that they aren't claiming to see the exact shape of the AI's brain. Because they only had 197 dots to work with in a huge 384-dimensional space, their "mesh" is an approximation, not a perfect reconstruction. They admit that with so few points, the "shape" they see is a topological approximation, not the intrinsic truth. However, the trends they observed are consistent and meaningful.

They explicitly rule out the idea that the AI is just collapsing into a trivial, useless point. The fact that some loops ( $\beta_1$ ) survived the whole training process proves the structure remained complex. They also rule out the idea that the AI learns in a single, sudden jump; instead, the changes are continuous, with the layers becoming more similar to each other as the training matures.

The Road Ahead

This paper is just the fourth stop on a long journey. The researchers have a plan for the future:

TGO-V (Computational Geometry): They want to know how the AI's math (the dot products and attention mechanisms) actually creates these shapes.
TGO-VI (Optimization Dynamics): They want to understand the "physics" of the learning process—how the math of the training algorithm (gradients and loss landscapes) forces the AI to settle into these stable shapes.

In short, TGO-IV gives us a new pair of glasses. Instead of just seeing the AI's data as a cloud of numbers, we can now see it as a living, breathing shape that starts as a scattered archipelago and grows into a stable, connected continent. It suggests that when a Transformer learns, it isn't just memorizing; it is building a stable, topological home for its knowledge.

Technical Summary: TGO-IV: Developmental Topology Observatory

1. Problem Statement and Motivation

While Transformers have demonstrated profound success in language processing and computer vision, the mechanisms underlying their learning dynamics remain partially opaque. Existing interpretability studies, such as those utilizing CKA, SVCCA, or spectral analyses (e.g., TGO-I, TGO-II, TGO-III), primarily focus on isolated layers, loss landscapes, or local geometric properties like covariance spectra and intrinsic dimensionality.

A critical gap remains in understanding the global topological structure of representation manifolds as they evolve through the forward pass. Geometric analyses are inherently insensitive to global topological invariants; two manifolds can share identical local geometry while possessing vastly different global connectivity (e.g., holes, voids, or connected components). The paper posits that current methods fail to determine whether Transformer layers preserve, create, merge, or eliminate fundamental topological structures during representation development. TGO-IV addresses this by introducing a topological framework to analyze the developmental evolution of Transformer representations.

2. Methodology

TGO-IV employs Topological Data Analysis (TDA), specifically Persistent Homology, to study the evolution of token-level representation point clouds across Transformer layers. Unlike visualization techniques (PCA, t-SNE, UMAP) that project data into lower dimensions and distort metric relationships, TGO-IV performs all analyses directly within the original high-dimensional embedding space ( $R^d$ ).

Core Pipeline

Point Cloud Construction: For a Vision Transformer (ViT), the token embeddings at each layer $l$ are treated as a point cloud $X^{(l)} = \{x^{(l)}_1, \dots, x^{(l)}_N\}$ in $R^d$ .
Simplicial Complex Construction: The paper constructs Vietoris–Rips simplicial complexes from these point clouds. Given a filtration radius $\epsilon$ , points are connected if their pairwise distance $\|x_i - x_j\| \le \epsilon$ .
Filtration and Persistence: A nested family of complexes is generated by increasing $\epsilon$ . Persistent Homology tracks the "birth" and "death" of topological features (connected components, loops, voids) as the filtration scale changes.
Topological Observatories: The framework utilizes a suite of complementary metrics to quantify the evolution of these structures:
- Persistence Diagrams & Barcode Diagrams: Visualize the lifetime of topological features.
- Betti Curves: Quantify the number of features ( $\beta_0$ for components, $\beta_1$ for loops, $\beta_2$ for voids) at each filtration scale.
- Persistence Landscapes: Transform diagrams into functional representations for statistical comparison.
- Bottleneck & Wasserstein Distances: Measure the dissimilarity between topological signatures of consecutive layers or training stages.

Experimental Setup

Model: ViT-Small/16 trained on ImageNet-100 for 100 epochs.
Data: A fixed validation subset of 1,000 images and a single designated image for detailed tracking.
Constraint: The authors explicitly note that with 197 tokens in a 384-dimensional space, the sampling is sparse. Therefore, the constructed complexes are treated as topological approximations rather than faithful reconstructions of the true intrinsic manifold. The goal is to identify consistent developmental trends in these signatures, not to claim exact manifold topology.

3. Key Contributions

The paper outlines four primary contributions:

First Topological Observatory: Introduces the first framework to apply Persistent Homology to Transformer representations, extending the TGO series from spectral and geometric analysis to topological data analysis.
Topology-Centric Framework: Formulates a method to construct Vietoris–Rips complexes from token-level point clouds and systematically characterize their evolution layer-by-layer.
Complementary Observatory Suite: Proposes a comprehensive set of metrics (Persistence Diagrams, Barcodes, Betti Curves, Landscapes, Bottleneck, and Wasserstein distances) to quantify the birth, persistence, and disappearance of topological features.
Developmental Evolution Analysis: Investigates how persistent topological signatures emerge, transform, and stabilize throughout the network, offering insights beyond conventional geometric and spectral analyses.

4. Key Results and Findings

The empirical analysis of the ViT-Small/16 training process revealed several consistent developmental trends:

Progressive Topological Stabilization: The Bottleneck Distance and Wasserstein Distance between consecutive layers decreased significantly as training progressed. This indicates that the persistent topological signatures of adjacent layers become increasingly similar as the model converges.
Connectivity Evolution (Betti Curves): The zeroth Betti number ( $\beta_0$ ), representing connected components, dominated the topology. As training advanced, the $\beta_0$ curves shifted toward smaller filtration radii, meaning connected components merged earlier. This suggests the representation point cloud becomes progressively more cohesive.
Persistence of Loops: The first Betti number ( $\beta_1$ ), representing one-dimensional loops, remained sparse but non-zero throughout training. This argues against a "trivial collapse" of the representation into a featureless point cloud; non-trivial geometric organization is preserved.
Absence of Higher-Order Cavities: The second Betti number ( $\beta_2$ ) remained nearly zero, indicating that higher-dimensional voids are not a significant component of the persistent topology in this setting.
Layer-Specific Transitions: While global similarity increased, localized peaks in topological distance persisted around specific layers (e.g., Layers 3, 10, and 12), suggesting these stages perform more pronounced topological reorganizations even after convergence.

5. Hypothesis and Significance

Based on the observed trends, the authors propose the Progressive Topological Stabilization Hypothesis. This hypothesis posits that Transformer representations evolve through a process where the point cloud becomes increasingly compact while preserving non-trivial geometric organization. During optimization, higher-order topological structures are transiently established and subsequently refined, while successive layers perform progressively smaller cumulative topological transformations as convergence is approached.

Significance Claims:

Beyond Geometry: The paper claims to provide a perspective that geometric and spectral analyses cannot offer: the ability to detect global topological transitions (e.g., merging of components vs. creation of holes) that are invisible to local metrics.
Mechanistic Insight: It moves the field closer to answering "how" Transformers learn by characterizing the developmental evolution of representations rather than just their static state.
Foundation for Future Work: TGO-IV serves as a foundational step for the broader Transformer Geometry Observatory (TGO) framework. The authors frame this work as a descriptive step that sets the stage for future observatories (TGO-V and TGO-VI) which aim to explain the computational mechanisms (attention routing, gradient dynamics) and optimization principles responsible for these topological changes.

The paper maintains a modest tone, emphasizing that these findings describe the evolution of constructed simplicial complexes based on sparse sampling, rather than definitive claims about the intrinsic topology of the underlying representation manifold.