University of Southampton
Towards understanding geometric properties in representation learning
Abstract
dc:description.abstractRepresentation learning seeks to discover structured, compact descriptions of data that support downstream reasoning. Central to this goal is the question of how geometric information, including both the spatial structure of inputs and the transformations that act upon them, is encoded, preserved, and measured within learned representations. This thesis investigates these properties from three complementary perspectives. First, Three-Dimensional Krawtchouk Moment Invariants (3DKMI) are developed as a provably invariant, training-free approach to volumetric shape description, demonstrating that mathematically guaranteed invariance translates into practical discriminative power for species classification. Second, a systematic comparison of segmentation architectures for 3D chamber delineation shows that the choice of supervision target actively sculpts the geometry of learned representations, with joint interior and boundary learning improving spatial precision and enabling the first automated reconstruction of individual growth trajectories. Third, SEIS (Subspace-based Equivariance and Invariance Scores) is introduced as a subspace-based framework that disentangles equivariance from invariance in neural representations across layers and architectures, revealing that these properties are not in strict competition but are simultaneously shaped by training decisions including data augmentation, task objectives, and architectural design. Together, these contributions establish that geometric properties in representation learning are actively negotiated through the interplay of construction, supervision, and learning, with broad implications for how spatial structure is embedded, controlled, and organized in imaging and vision systems.
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Lin, Huahua
- Advisors dc:contributor.advisor
-
- Cai, Xiaohao
- Farrahi, Kate
- Ezard, Tom
- ORCID person:identifier.orcid
Rights
- Language dc:language
- English, eng
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5258/SOTON/01/T231