What is Common Knowledge Across Domains? Investigating Shared Representations in Transfer Learning

Author

• Parth Ganeriwala, Florida Institute of TechnologyFollow

Date of Award

5-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

First Advisor

Siddhartha Bhattacharyya

Second Advisor

Ryan T. White

Third Advisor

Eraldo Ribeiro

Fourth Advisor

Randolph M. Jones

Abstract

Transfer learning (TL) has become an essential approach for extending machine learning (ML) models by utilizing existing knowledge across different domains. Understanding and leveraging common knowledge—shared representations across various domains—is critical for optimizing TL. Yet TL research rarely specifies what is being transferred in a way that is both theoretically meaningful and experimentally testable. This research work addresses that gap by formalizing common knowledge as a shared representational subspace: a set of features and transformations that remains useful, stable, and causally important across domains and tasks, and that a model can reuse without domain-specific modification. This work operationalizes the definition with a convergent-evidence framework integrating representational similarity, transfer under controlled freezing and low-supervision regimes, catastrophic-forgetting measurements, cross-domain probing of geometric primitives, and layerwise visualization of learned features. The framework is evaluated through a cross-domain case study in visual lane and marking interpretation, of transfer from road lane datasets to aircraft taxiway and runway markings.  This work also introduces AssistTaxi, a real-world taxiway dataset, and a hybrid labeling methodology that reduces annotation effort while preserving safety-critical label fidelity. Representational similarity analyses, measured using Centered Kernel Alignment (CKA, scaled to [0, 1] where 1 indicates identical representations), indicated approximately 0.7 similarity in early backbone layers across CuLane and CurveLanes source models, with similarity decreasing toward later layers. Freezing early backbone layers and fine-tuning on AssistTaxi improved performance relative to training from scratch, and interpretability analyses associated the shared subspace with geometric line cues. The analyses characterize when transfer is beneficial, where it fails, and which representational components behave as stable, reusable knowledge versus domain-specific specialization. By defining common knowledge as a measurable construct and providing a convergent-evidence methodology to identify it, this work clarifies the conditions under which TL generalizes reliably across domain shifts. The resulting insights support the development of more interpretable and adaptable perception models that leverage shared geometric structure while preserving domain-specific details necessary for high task performance.

Recommended Citation

Ganeriwala, Parth, "What is Common Knowledge Across Domains? Investigating Shared Representations in Transfer Learning" (2026). Theses and Dissertations. 1634.
https://repository.fit.edu/etd/1634