BOSE: Block-Wise Federated Learning in Heterogeneous Edge Computing

They divide na original DL model into several non-overlapping blocks, so they can be trained separately on the low-capability devices. Each block has its own output, therefore, Block-Wise Training is employed.

Several Evaluations in used in this system, which are designed to dynamically adjust the Block-Wise Trianing. At each round, the parameter server will adaptively assign blocks for devices to perform local training.

• Block Potential Evaluation.
• Device Resource Estimation.
• Round Deadline Adjustment.

Rather than evaluating block potential by measuring gradient magnitude, this system adopt the learning rate as an important indicator: \(P_b^k=\frac{||\sum^{r-1}_{i=0}∆^{k-i}_b||}{ε+\sum^{r-1}_{i=0}|| ∆^{k-i}_b||}\) \(∆^k_b=w^k_b-w^{k-1}_b,k>=2\)

$k$ is the updating of block b at round k and r is the observation window.