Energy Efficient And Fast Controlled Descent For Over The Air Assisted Federated Learning

Abstract: ABSTRACT ENERGY-EFFICIENT AND FAST CONTROLLED DESCENT FOR OVER-THE-AIR ASSISTED FEDERATED LEARNING Present disclosure relates to method and system for performing Over-the-Air Computation-Assisted Federated Learning (AEFL). Method utilizes an energy-efficient controlled decent approach that adapts receiver’s (e.g., edge devices) transmit power and transmitter’s (e.g., parameter server) amplification factor to control a fixed or variable rate of convergence of AEFL, and control the number of iterations and energy required to converge, while minimizing error floor contributed by each iteration. Method is based on recursive upper bound on optimality gap, which measures the performance of AEFL technique. It is difference between a loss function given a current local model computed over datasets of all the receivers in an iteration and its minimum value if the data were centrally available at the transmitter. In every iteration, optimality gap contracts by a factor that depends on an effective learning rate, which is different from learning rate specified in min-MSE scheme, and increases by an error floor. Fig. 1B

Patent Information

Applicants

Inventors

Specification

Documents

Application Documents

Search Strategy

ERegister / Renewals Inforce