IoT Seminar | Synergizing Federated Learning and Foundational Models: Bidirectional Enhancements for Collaborative Training and Knowledge Transfer

Federated learning traditionally operates on small, decentralized data collected from resource-limited edge devices, while large language models (LLMs) are trained on massive centralized datasets with extensive computational resources. This fundamental contrast raises a major challenge: how to harness the generalization power of LLMs within the privacy-preserving, decentralized framework of federated learning. To address this, I developed GPT-FL, a federated learning framework that integrates foundational models into edge-device training pipelines. GPT-FL enables lightweight edge models to benefit from the capabilities of centralized LLMs through prompt-based distillation and task decomposition, improving downstream performance while preserving user privacy, accommodating heterogeneity, and reducing communication overhead. Building upon this foundation, I further explore how federated learning inspires new paradigms for federated LLMs, including the design of uncertainty-guided routers to optimize inference across model families under cost and performance constraints. Finally, I bring these academic contributions into real-world deployment by collaborating with the USC Pacific Asia Museum, building the first social-scale AI testbed that directly translates research into societal impact through live engagement with public audiences.