Adaptive stepsize estimation based accelerated gradient descent algorithm for fully complex-valued neural networks

Weijing Zhao, He Huang

Abstract: Highlights•Adaptive stepsize design methods without manual tuning are proposed for CNAG.•Stepsize is obtained by estimating the norm of approximate Hessian matrix.•Theoretical analysis is presented to support the validity of the design method.•Scaling parameter is computed by means of the importance of curvature information.•Efficient training of CVNNs is achieved by the proposed algorithms.