Solution of Nonconvex Nonsmooth Stochastic Optimization Problems

Different classes of nonconvex nonsmooth stochastic optimization problems are analyzed, their generalized differentiability properties and necessary optimality conditions are studied, and a technique for calculating stochastic gradients is developed. For each class of the problems, corresponding solution methods are proposed, in particular, generalizations of the stochastic quasigradient method.     

An analysis of a class of neural networks for solving linearprogramming problems

A class of neural networks that solve linear programming problems is analyzed. The neural networks considered are modeled by dynamic gradient systems that are constructed using a parametric family of exact (nondifferentiable) penalty functions. It is proved that for a given linear programming problem and sufficiently large penalty parameters, any trajectory of the neural network converges in finite time to its solution set. For the analysis, Lyapunov-type theorems are developed for finite time convergence of nonsmooth sliding mode dynamic systems to invariant sets. The results are illustrated via numerical simulation examples     

Almost periodic dynamics of the delayed complex-valued recurrent neural networks with discontinuous activation functions

The target of this article is to study almost periodic dynamical behaviors for complex-valued recurrent neural networks with discontinuous activation functions and time-varying delays. We construct an equivalent discontinuous right-hand equation by decomposing real and imaginary parts of complex-valued neural networks. Based on differential inclusions theory, diagonal dominant principle and nonsmooth analysis theory of generalized Lyapunov function method, we achieve the existence, uniqueness and global stability of almost periodic solution for the equivalent delayed differential network. In particular, we derive a series of results on the equivalent neural networks with discontinuous activation functions, constant coefficients as well as periodic coefficients, respectively. Finally, we give a numerical example to demonstrate the effectiveness and feasibility of the derived theoretical results.

     

Lyapunov stability and generalized invariance principle for nonconvex differential inclusions

This paper studies the system stability problems of a class of nonconvex differential inclusions. At first, a basic stability result is obtained by virtue of locally Lipschitz continuous Lyapunov functions. Moreover, a generalized invariance principle and related attraction conditions are proposed and proved to overcome the technical difficulties due to the absence of convexity. In the technical analysis, a novel set-valued derivative is proposed to deal with nonsmooth systems and nonsmooth Lyapunov functions. Additionally, the obtained results are consistent with the existing ones in the case of convex differential inclusions with regular Lyapunov functions. Finally, illustrative examples are given to show the effectiveness of the methods.     

Convergence of Neural Networks for Programming Problems via a Nonsmooth Łojasiewicz Inequality

This paper considers a class of neural networks (NNs) for solving linear programming (LP) problems, convex quadratic programming (QP) problems, and nonconvex QP problems where an indefinite quadratic objective function is subject to a set of affine constraints. The NNs are characterized by constraint neurons modeled by ideal diodes with vertical segments in their characteristic, which enable to implement an exact penalty method. A new method is exploited to address convergence of trajectories, which is based on a nonsmooth Lstrokojasiewicz inequality for the generalized gradient vector field describing the NN dynamics. The method permits to prove that each forward trajectory of the NN has finite length, and as a consequence it converges toward a singleton. Furthermore, by means of a quantitative evaluation of the Lstrokojasiewicz exponent at the equilibrium points, the following results on convergence rate of trajectories are established: 1) for nonconvex QP problems, each trajectory is either exponentially convergent, or convergent in finite time, toward a singleton belonging to the set of constrained critical points; 2) for convex QP problems, the same result as in 1) holds; moreover, the singleton belongs to the set of global minimizers; and 3) for LP problems, each trajectory converges in finite time to a singleton belonging to the set of global minimizers. These results, which improve previous results obtained via the Lyapunov approach, are true independently of the nature of the set of equilibrium points, and in particular they hold even when the NN possesses infinitely many nonisolated equilibrium points     

ADMM for multiaffine constrained optimization

ABSTRACT

We expand the scope of the alternating direction method of multipliers (ADMM). Specifically, we show that ADMM, when employed to solve problems with multiaffine constraints that satisfy certain verifiable assumptions, converges to the set of constrained stationary points if the penalty parameter in the augmented Lagrangian is sufficiently large. When the Kurdyka–?ojasiewicz (K–?) property holds, this is strengthened to convergence to a single constrained stationary point. Our analysis applies under assumptions that we have endeavoured to make as weak as possible. It applies to problems that involve nonconvex and/or nonsmooth objective terms, in addition to the multiaffine constraints that can involve multiple (three or more) blocks of variables. To illustrate the applicability of our results, we describe examples including nonnegative matrix factorization, sparse learning, risk parity portfolio selection, nonconvex formulations of convex problems and neural network training. In each case, our ADMM approach encounters only subproblems that have closed-form solutions.     

Representation of Nonlinear Random Transformations by Non-Gaussian Stochastic Neural Networks

The learning capability of neural networks is equivalent to modeling physical events that occur in the real environment. Several early works have demonstrated that neural networks belonging to some classes are universal approximators of input-output deterministic functions. Recent works extend the ability of neural networks in approximating random functions using a class of networks named stochastic neural networks (SNN). In the language of system theory, the approximation of both deterministic and stochastic functions falls within the identification of nonlinear no-memory systems. However, all the results presented so far are restricted to the case of Gaussian stochastic processes (SPs) only, or to linear transformations that guarantee this property. This paper aims at investigating the ability of stochastic neural networks to approximate nonlinear input-output random transformations, thus widening the range of applicability of these networks to nonlinear systems with memory. In particular, this study shows that networks belonging to a class named non-Gaussian stochastic approximate identity neural networks (SAINNs) are capable of approximating the solutions of large classes of nonlinear random ordinary differential transformations. The effectiveness of this approach is demonstrated and discussed by some application examples.     

Trust-region algorithms for training responses: machine learning methods using indefinite Hessian approximations

ABSTRACT

Machine learning (ML) problems are often posed as highly nonlinear and nonconvex unconstrained optimization problems. Methods for solving ML problems based on stochastic gradient descent are easily scaled for very large problems but may involve fine-tuning many hyper-parameters. Quasi-Newton approaches based on the limited-memory Broyden-Fletcher-Goldfarb-Shanno (BFGS) update typically do not require manually tuning hyper-parameters but suffer from approximating a potentially indefinite Hessian with a positive-definite matrix. Hessian-free methods leverage the ability to perform Hessian-vector multiplication without needing the entire Hessian matrix, but each iteration's complexity is significantly greater than quasi-Newton methods. In this paper we propose an alternative approach for solving ML problems based on a quasi-Newton trust-region framework for solving large-scale optimization problems that allow for indefinite Hessian approximations. Numerical experiments on a standard testing data set show that with a fixed computational time budget, the proposed methods achieve better results than the traditional limited-memory BFGS and the Hessian-free methods.     

Lyapunov stability theory of nonsmooth systems

This paper develops nonsmooth Lyapunov stability theory and LaSalle's invariance principle for a class of nonsmooth Lipschitz continuous Lyapunov functions and absolutely continuous state trajectories. Computable tests based on Filipov's differential inclusion and Clarke's generalized gradient are derived. The primary use of these results is in analyzing the stability of equilibria of differential equations with discontinuous right-hand side such as in nonsmooth dynamic systems or variable structure control     

Robust-nonsmooth Kalman Filtering for Stochastic Sandwich Systems with Dead-zone

In this paper, a robust-nonsmooth Kalman filtering approach for stochastic sandwich systems with dead-zone is proposed, which can guarantee the variance of filtering error to be upper bounded. In this approach, the stochastic sandwich system with dead-zone is described by a stochastic nonsmooth state-space function. Then, in order to approximate the nonsmooth sandwich system within a bounded region around the equilibrium point, a linearization approach based on nonsmooth optimization is proposed. For handling the model uncertainty caused by linearization and modeling, the robust-nonsmooth Kalman filtering method is proposed for state estimation of the stochastic sandwich system with dead-zones with model uncertainty. Finally, both simulation and experimental examples are presented for evaluating the performance of the proposed filtering scheme.

     

Nonsmooth computational mechanics algorithms, quasidifferentiability and related topics

Several extensions of smooth computational mechanics algorithms for the treatment of nonsmooth and possible nonconvex problems are briefly discussed in this paper. A potential or dissipation energy minimization problem approach is used for the structural analysis problem, so as to make the link with mathematical optimization techniques straightforward. Variational inequality problems, hemivariational inequality problems and systems of variational inequalities can be treated by the methods reviewed in this paper. The use of quasidifferentiable and codifferentiable optimization techniques is proposed for the solution of the more general class of nonconvex, possibly nonsmooth problems. Established and new directions in path-following techniques are discussed and are linked with nonsmooth mechanics algorithms.     

基于ArcReLU函数的神经网络激活函数优化研究

近年来深度学习发展迅猛。由于深度学习的概念源于神经网络,而激活函数更是神经网络模型在学习理解非线性函数时不可或缺的部分,因此本文对常用的激活函数进行了研究比较。针对常用的激活函数在反向传播神经网络中具有收敛速度较慢、存在局部极小或梯度消失的问题,将Sigmoid系和ReLU系激活函数进行了对比,分别讨论了其性能,详细分析了几类常用激活函数的优点及不足,并通过研究Arctan函数在神经网络中应用的可能性,结合ReLU函数,提出了一种新型的激活函数ArcReLU。实验证明,该函数既能显著加快反向传播神经网络的训练速度,又能有效降低训练误差并避免梯度消失的问题。     

神经网络增强学习的梯度算法研究

针对具有连续状态和离散行为空间的Markov决策问题，提出了一种新的采用多层前馈神经网络进行值函数逼近的梯度下降增强学习算法，该算法采用了近似贪心且连续可微的Boltzmann分布行为选择策略，通过极小化具有非平稳行为策略的Bellman残差平方和性能指标，以实现对Markov决策过程最优值函数的逼近，对算法的收敛性和近似最优策略的性能进行了理论分析，通过Mountain-Car学习控制问题的仿真研究进一步验证了算法的学习效率和泛化性能。     

Inexact proximal stochastic second-order methods for nonconvex composite optimization

ABSTRACT

In this paper, we propose a framework of Inexact Proximal Stochastic Second-order (IPSS) method for solving nonconvex optimization problems, whose objective function consists of an average of finitely many, possibly weakly, smooth functions and a convex but possibly nonsmooth function. At each iteration, IPSS inexactly solves a proximal subproblem constructed by using some positive definite matrix which could capture the second-order information of original problem. Proper tolerances are given for the subproblem solution in order to maintain global convergence and the desired overall complexity of the algorithm. Under mild conditions, we analyse the computational complexity related to the evaluations on the component gradient of the smooth function. We also investigate the number of evaluations of subgradient when using an iterative subgradient method to solve the subproblem. In addition, based on IPSS, we propose a linearly convergent algorithm under the proximal Polyak–?ojasiewicz condition. Finally, we extend the analysis to problems with weakly smooth function and obtain the computational complexity accordingly.     

Fusion of learning automata theory and granular inference systems: ANLAGIS. Applications to pattern recognition and machine learning

In this paper the fusion of artificial neural networks, granular computing and learning automata theory is proposed and we present as a final result ANLAGIS, an adaptive neuron-like network based on learning automata and granular inference systems. ANLAGIS can be applied to both pattern recognition and learning control problems. Another interesting contribution of this paper is the distinction between pre-synaptic and post-synaptic learning in artificial neural networks. To illustrate the capabilities of ANLAGIS some experiments on knowledge discovery in data mining and machine learning are presented. The main, novel contribution of ANLAGIS is the incorporation of Learning Automata Theory within its structure; the paper includes also a novel learning scheme for stochastic learning automata.     

深度学习应用技术研究

本文针对深度学习应用技术进行了研究性综述。详细阐述了RBM（Restricted Boltzmann Machine）逐层预训练后再用BP（back-propagation）微调的深度学习贪婪层训练方法,对比分析了BP算法中三种梯度下降的方式,建议在线学习系统,采用随机梯度下降,静态离线学习系统采用随机小批量梯度下降;归纳总结了深度学习深层结构特征,并推荐了目前最受欢迎的5层深度网络结构设计方法。分析了前馈神经网络非线性激活函数的必要性及常用的激活函数优点,并推荐ReLU （rectified linear units）激活函数。最后简要概括了深度CNNs(Convolutional Neural Networks), 深度RNNs(recurrent neural networks), LSTM(long short-termmemory networks)等新型深度网络的特点及应用场景,并归纳总结了当前深度学习可能的发展方向。     

采用分数阶动量的卷积神经网络随机梯度下降法

针对随机梯度下降法可能会收敛到局部最优的问题,文中提出采用分数阶动量的随机梯度下降法,提高卷积神经网络的识别精度和学习收敛速度.结合基于动量的随机梯度下降法和分数阶差分运算,改进参数更新方法,讨论分数阶阶次对网络参数训练效果的影响,给出阶次调整方法.在MNIST、CIFAR-10数据集上的实验表明,文中方法可以提高卷积神经网络的识别精度和学习收敛速度.     

Stochastic competitive learning

Competitive learning systems are examined as stochastic dynamical systems. This includes continuous and discrete formulations of unsupervised, supervised, and differential competitive learning systems. These systems estimate an unknown probability density function from random pattern samples and behave as adaptive vector quantizers. Synaptic vectors, in feedforward competitive neural networks, quantize the pattern space and converge to pattern class centroids or local probability maxima. A stochastic Lyapunov argument shows that competitive synaptic vectors converge to centroids exponentially quickly and reduces competitive learning to stochastic gradient descent. Convergence does not depend on a specific dynamical model of how neuronal activations change. These results extend to competitive estimation of local covariances and higher order statistics.     

Deep learning: an overview and main paradigms

In the present paper, we examine and analyze main paradigms of learning of multilayer neural networks starting with a single layer perceptron and ending with deep neural networks, which are considered regarded as a breakthrough in the field of the intelligent data processing. The baselessness of some ideas about the capacity of multilayer neural networks is shown and transition to deep neural networks is justified. We discuss the principal learning models of deep neural networks based on the restricted Boltzmann machine (RBM), an autoassociative approach and a stochastic gradient method with a Rectified Linear Unit (ReLU) activation function of neural elements.     

机器学习随机优化方法的个体收敛性研究综述

随机优化方法是求解大规模机器学习问题的主流方法,其研究的焦点问题是算法是否达到最优收敛速率与能否保证学习问题的结构。目前,正则化损失函数问题已得到了众多形式的随机优化算法,但绝大多数只是对迭代进行 平均的输出方式讨论了收敛速率,甚至无法保证最为典型的稀疏结构。与之不同的是,个体解能很好保持稀疏性,其最优收敛速率已经作为open问题被广泛探索。另外,随机优化普遍采用的梯度无偏假设往往不成立,加速方法收敛界中的偏差在有偏情形下会随迭代累积,从而无法应用。本文对一阶随机梯度方法的研究现状及存在的问题进行综述,其中包括个体收敛速率、梯度有偏情形以及非凸优化问题,并在此基础上指出了一些值得研究的问题。