Morphing Distributed and Transfer Learning Paradigms for Visual Understanding

This thesis addresses key limitations of machine learning (ML) models for visual understanding, particularly their difficulty in adapting to dynamic, real-world conditions without requiring extensive retraining. We focus on three emerging paradigms, Domain Adaptation, Continual Learning, and Federated Learning, to tackle this challenge. Domain Adaptation enables models to adapt to new data distributions, Continual Learning allows systems to learn from a continuous stream of information without forgetting prior knowledge, and Federated Learning offers a privacy-preserving approach to model training by learning from decentralized data sources without the need for raw data centralization. By revisiting and combining these paradigms, this thesis seeks to propose novel insights into visual understanding, addressing practical challenges related to model adaptability, efficiency, and privacy. The insights gained from this research contribute to the development of ML systems that are more flexible, generalize better across domains, and respect user privacy, all while reducing computational and environmental costs.