Adaptive Subcarrier, Parameter, and Power Allocation for Partitioned Edge Learning Over Broadband Channels

In this paper, we consider partitioned edge learning (PARTEL), which implements parameter-server training, a well known distributed learning method, in wireless network. Thereby, PARTEL leverages computation resources at devices to train large-scale artificial intelligence (AI) model by dynamically partitioning the into parametric blocks for separated updating devices. Targeting broadband channels, joint control of parameter allocation, sub-channel and transmission power improve performance PARTEL. Specifically, policies SUbcarrier, Parameter, POweR allocaTion (SUPPORT) are optimized under criterion minimum latency. Two cases considered. First, case decomposable models (e.g., logistic regression), latency-minimization problem is mixed-integer program non-convex. Due its intractability, develop practical solution integer relaxation transforming it an equivalent convex size maximization latency constraint. low-complexity algorithm designed compute SUPPORT policy. Second, deep neural network (DNN) can be trained using introducing some auxiliary variables. This, however, introduces constraints on reducing granularity allocation. The preceding policy extended DNN applying proposed techniques load rounding proportional adjustment rein expansion caused constraints.