Federated learning often needs to deliver updated models within tight deadlines, especially when data distributions shift over time. The new TALLOP method addresses this by jointly optimizing how much local training each client performs and how aggressively model updates are compressed before transmission.
Rather than treating local computation and communication as separate tuning problems, TALLOP coordinates them under a wall-clock budget. This allows the system to balance the cost of extra local steps against the bandwidth savings from compression, with the goal of hitting a target accuracy as quickly as possible.
The paper emphasizes heterogeneous environments, where devices differ in capability and network conditions. By adapting both training and compression decisions in a unified framework, TALLOP aims to make time-constrained federated learning more practical in real-world deployments.