Professor Yong Liu's general research area is networking. His current research projects include resilient edge networks, immersive video streaming for AR/VR applications, and decentralized federated learning (DFL) and applications. For this research visit, Professor Liu is interested in exploring research collaborations on all the above topics with faculty and other fellows in NYUSH, in particular on DFL. Mainstream machine learning operates in a centralized “walled garden”, with data, model, and computing resources collocated within data centers. Not only does this create increasing entry barriers, but also raises privacy concerns among users about how their data were used by the few big players to train their private AI models. Decentralized Federated Learning (DFL) addresses this by enabling a swarm of agents to collaboratively train a shared model without sharing private data. Each agent processes its multi-modal data locally, exchanging only model updates via peer-to-peer communication to reach a consensus global model. However, DFL faces significant challenges due to resource-limited agents and best-effort Internet connections. Knowledge representation, dissemination, and fusion become complex, demanding excessive communication and computation. Professor Liu's research holistically addresses these challenges through integrated Communication, Computing, and Control (3C) designs. Key investigations include:
1. Adaptive overlay topology control for fast and resilient DFL convergence in the face of dynamic and heterogeneous agents. Energy-efficient multimodal DFL that flexibly
partitions multimodal data training among agents, and adaptively adjusts learning speed and model fusion strategy.
2. Delay-tolerant knowledge dissemination on mobile agents through model caching. Investigating the trade off between model staleness and coverage. Develop novel multimodal model caching algorithms for agents with different mobility patterns.
3. Adaptive multimodal knowledge streaming and trade-off between communication overhead of streaming and computation overhead of modality translation.
4. Applications of DFL in Connected Autonomous Vehicles (CAVs) that leverage on data collected by distributed vehicles for DFL-based realtime decision making.