Transformer Geometry Observatory TGO-II: Representational Similarity Observatory
2026-07-02 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionMachine Learning
AI summaryⓘ
The authors study how Vision Transformers change their internal data representations as they learn. They find that layers become more specialized and the complexity of the data representation increases during training. Interestingly, the way tokens (pieces of input data) interact remains strong rather than becoming more independent. This means the model's complexity grows by richer transformations while keeping token relationships intact. Their work offers insights into how these models learn beyond just looking at performance or attention.
Vision TransformerRepresentation GeometryCentered Kernel Alignment (CKA)Singular Vector Canonical Correlation Analysis (SVCCA)Intrinsic DimensionalityToken CovarianceManifold ExpansionToken InteractionSupervised Training
Authors
Kaustubh Kapil, Kishor P. Upla
Abstract
While Vision Transformers have achieved remarkable success across computer vision and language applications, the geometric evolution of their internal representations throughout training remains insufficiently understood. Existing analyses primarily focus on attention mechanisms and downstream performance, leaving the evolution of representation geometry largely unexplored. In this work, we present Transformer Geometry Observatory-II (TGO-II), a representation geometry analysis framework designed to investigate how Transformer representations evolve during supervised training. TGO-II analyzes Vision Transformer (ViT-Small/16) representations using Centered Kernel Alignment (CKA), Singular Vector Canonical Correlation Analysis (SVCCA), Two-Nearest Neighbor Intrinsic Dimensionality (TwoNN-ID), and token covariance analysis. Our experiments reveal three key observations. First, both CKA and SVCCA progressively decrease throughout training, indicating increasing representational specialization across Transformer layers. Second, intrinsic dimensionality consistently increases before stabilizing, suggesting progressive expansion of the representation manifold into a larger set of locally accessible degrees of freedom. Third, token covariance and coupling analyses demonstrate that strong token interaction structure persists throughout training, challenging the hypothesis that increasing representational complexity arises primarily from progressive token independence. These findings suggest that representation complexity and layer specialization emerge simultaneously during training. Manifold expansion appears to occur without token decoupling. Together, these observations motivate a new hypothesis in which Vision Transformers increase representational complexity through progressively richer transformations while preserving strong token interaction structure during learning.