CollEdge: Collective Intelligence Emerges on Edge Learning Systems with Privacy-Preserving Knowledge Sharing and Lifelong Evolution

The rapid proliferation of edge devices with embedded AI capabilities enables local processing of private data, providing an inherent foundation for privacy protection. However, most current edge learning systems still depend on centralized aggregation, where updated local AI models are periodically transmitted to a server. This approach introduces inefficiencies in training and undermines robustness due to single-point failures. To address these limitations, CollEdge will design and implement a decentralized, privacy-preserving AI paradigm that fosters the emergence of collective intelligence on edge devices.CollEdge will empower edge clients to evolve continually while enabling collaborative knowledge sharing: isolated devices will self-organize through decentralized coordination mechanisms that safeguard privacy, allowing system-level intelligence to emerge as a significant property of collaboration. First, edge devices will be equipped with lifelong learning capabilities to retain essential past knowledge and integrate newly coming information under hardware and resource constraints. Then, collective knowledge sharing will be realized through decentralized federated learning, ensuring privacy without raw data exchange. Finally, a co-optimization strategy spanning computation and network topology will unlock the full potential of large-scale multi-device systems. The paradigm will be validated in three domains of economic, strategic, and environmental importance: healthcare, embodied AI, and battery state estimation.CollEdge will lead edge devices to form a more capable collective than isolated clients by ensuring lifelong evolution and privacy preservation. This fellowship combines tailored training activities at EPFL with a secondment to Oxford University, equipping me with the expertise and independence to advance research at the intersection of AI, healthcare, and information privacy.