🤖 AI Summary
This study addresses the lack of systematic analysis regarding the geometric evolution of internal representations during Vision Transformer (ViT) training. The authors propose the TGO-II framework, which integrates Centered Kernel Alignment (CKA), Singular Vector Canonical Correlation Analysis (SVCCA), TwoNN intrinsic dimension estimation, and token covariance analysis. Applying this framework to ViT-Small/16 under supervised training, they uncover a tripartite geometric evolution pattern: progressive layer-wise specialization, an initial rise followed by stabilization of intrinsic dimensionality, and the persistent presence of strong token interaction structures. These findings demonstrate that increased representational complexity co-occurs with layer specialization without requiring token decorrelation, thereby challenging the conventional assumption that complexity arises from token independence and highlighting ViT’s capacity to achieve rich representational transformations through sustained strong token interactions.
📝 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.