ADMM Training Algorithms for Residual Networks: Convergence, Complexity and Parallel Training

We design a series of serial and parallel proximal point (gradient) ADMMs for the fully connected residual networks (FCResNets) training problem by introducing auxiliary variables. Convergence of the proximal point version is proven based on a Kurdyka-Lojasiewicz (KL) property analysis framework, and we can ensure a locally R-linear or sublinear convergence rate depending on the different ranges of the Kurdyka-Lojasiewicz (KL) exponent, in which a necessary auxiliary function is constructed to realize our goal. Moreover, the advantages of the parallel implementation in terms of lower time complexity and less (per-node) memory consumption are analyzed theoretically. To the best of our knowledge, this is the first work analyzing the convergence, convergence rate, time complexity and (per-node) runtime memory requirement of the ADMM applied in the FCResNets training problem theoretically. Experiments are reported to show the high speed, better performance, robustness and potential in the deep network training tasks. Finally, we present the advantage and potential of our parallel training in large-scale problems.

Convergence Rates of Training Deep Neural Networks via Alternating Minimization Methods

Training deep neural networks (DNNs) is an important and challenging optimization problem in machine learning due to its non-convexity and non-separable structure. The alternating minimization (AM) approaches split the composition structure of DNNs and have drawn great interest in the deep learning and optimization communities. In this paper, we propose a unified framework for analyzing the convergence rate of AM-type network training methods. Our analysis are based on the $j$-step sufficient decrease conditions and the Kurdyka-Lojasiewicz (KL) property, which relaxes the requirement of designing descent algorithms. We show the detailed local convergence rate if the KL exponent $\theta$ varies in $[0,1)$. Moreover, the local R-linear convergence is discussed under a stronger $j$-step sufficient decrease condition.