循环神经网络研究综述

循环神经网络是神经网络序列模型的主要实现形式,近几年得到迅速发展,其是机器翻译、机器问题回答、序列视频分析的标准处理手段,也是对于手写体自动合成、语音处理和图像生成等问题的主流建模手段.鉴于此,循环神经网络的各分支按照网络结构进行详细分类,大致分为3大类:一是衍生循环神经网络,这类网络是基于基本RNNs模型的结构衍生变体,即对RNNs的内部结构进行修改;二是组合循环神经网络,这类网络将其他一些经典的网络模型或结构与第一类衍生循环神经网络进行组合,得到更好的模型效果,是一种非常有效的手段;三是混合循环神经网络,这类网络模型既有不同网络模型的组合,又在RNNs内部结构上进行修改,是同属于前两类网络分类的结构.为了更加深入地理解循环神经网络,进一步介绍与循环神经网络经常混为一谈的递归神经网络结构以及递归神经网络与循环神经网络的区别和联系.在详略描述上述模型的应用背景、网络结构以及模型变种后,对各个模型的特点进行总结和比较,并对循环神经网络模型进行展望和总结.     

递归最小二乘循环神经网络

针对循环神经网络(Recurrent neural networks, RNNs)一阶优化算法学习效率不高和二阶优化算法时空开销过大,提出一种新的迷你批递归最小二乘优化算法.所提算法采用非激活线性输出误差替代传统的激活输出误差反向传播,并结合加权线性最小二乘目标函数关于隐藏层线性输出的等效梯度,逐层导出RNNs参数的迷你批递归最小二乘解.相较随机梯度下降算法,所提算法只在RNNs的隐藏层和输出层分别增加了一个协方差矩阵,其时间复杂度和空间复杂度仅为随机梯度下降算法的3倍左右.此外,本文还就所提算法的遗忘因子自适应问题和过拟合问题分别给出一种解决办法.仿真结果表明,无论是对序列数据的分类问题还是预测问题,所提算法的收敛速度要优于现有主流一阶优化算法,而且在超参数的设置上具有较好的鲁棒性.     

Markovian architectural bias of recurrent neural networks

In this paper, we elaborate upon the claim that clustering in the recurrent layer of recurrent neural networks (RNNs) reflects meaningful information processing states even prior to training. By concentrating on activation clusters in RNNs, while not throwing away the continuous state space network dynamics, we extract predictive models that we call neural prediction machines (NPMs). When RNNs with sigmoid activation functions are initialized with small weights (a common technique in the RNN community), the clusters of recurrent activations emerging prior to training are indeed meaningful and correspond to Markov prediction contexts. In this case, the extracted NPMs correspond to a class of Markov models, called variable memory length Markov models (VLMMs). In order to appreciate how much information has really been induced during the training, the RNN performance should always be compared with that of VLMMs and NPMs extracted before training as the "" base models. Our arguments are supported by experiments on a chaotic symbolic sequence and a context-free language with a deep recursive structure.     

Multiobjective hybrid optimization and training of recurrent neural networks.

The application of neural networks to solve a problem involves tasks with a high computational cost until a suitable network is found, and these tasks mainly involve the selection of the network topology and the training step. We usually select the network structure by means of a trial-and-error procedure, and we then train the network. In the case of recurrent neural networks (RNNs), the lack of suitable training algorithms sometimes hampers these procedures due to vanishing gradient problems. This paper addresses the simultaneous training and topology optimization of RNNs using multiobjective hybrid procedures. The proposal is based on the SPEA2 and NSGA2 algorithms for making hybrid methods using the Baldwinian hybridization strategy. We also study the effects of the selection of the objectives, crossover, and mutation in the diversity during evolution. The proposals are tested in the experimental section to train and optimize the networks in the competition on artificial time-series (CATS) benchmark.     

The copula echo state network

Echo state networks (ESNs) constitute a novel approach to recurrent neural network (RNN) training, with an RNN (the reservoir) being generated randomly, and only a readout being trained using a simple, computationally efficient algorithm. ESNs have greatly facilitated the practical application of RNNs, outperforming classical approaches on a number of benchmark tasks. This paper studies the formulation of a class of copula-based semiparametric models for sequential data modeling, characterized by nonparametric marginal distributions modeled by postulating suitable echo state networks, and parametric copula functions that help capture all the scale-free temporal dependence of the modeled processes. We provide a simple algorithm for the data-driven estimation of the marginal distribution and the copula parameters of our model under the maximum-likelihood framework. We exhibit the merits of our approach by considering a number of applications; as we show, our method offers a significant enhancement in the dynamical data modeling capabilities of ESNs, without significant compromises in the algorithm's computational efficiency.     

Stability analysis of discrete-time recurrent neural networks

We address the problem of global Lyapunov stability of discrete-time recurrent neural networks (RNNs) in the unforced (unperturbed) setting. It is assumed that network weights are fixed to some values, for example, those attained after training. Based on classical results of the theory of absolute stability, we propose a new approach for the stability analysis of RNNs with sector-type monotone nonlinearities and nonzero biases. We devise a simple state-space transformation to convert the original RNN equations to a form suitable for our stability analysis. We then present appropriate linear matrix inequalities (LMIs) to be solved to determine whether the system under study is globally exponentially stable. Unlike previous treatments, our approach readily permits one to account for non-zero biases usually present in RNNs for improved approximation capabilities. We show how recent results of others on the stability analysis of RNNs can be interpreted as special cases within our approach. We illustrate how to use our approach with examples. Though illustrated on the stability analysis of recurrent multilayer perceptrons, the approach proposed can also be applied to other forms of time-lagged RNNs.     

Combining GRN modeling and demonstration-based programming for robot control

In recent years, gene regulatory networks (GRNs) have been proposed to work as reliable and robust control mechanisms for robots. Because recurrent neural networks (RNNs) have the unique characteristic of presenting system dynamics over time, we thus adopt such kind of network structure and the principles of gene regulation to develop a biologically and computationally plausible GRN model for robot control. To simulate the regulatory effects and to make our model inferable from time-series data, we also implement an enhanced network-learning algorithm to derive network parameters efficiently. In addition, we present a procedure of programming-by-demonstration to collect behavior sequence data of the robot as expression profiles, and then employ our network-modeling framework to infer controllers. To verify the proposed approach, experiments have been conducted, and the results show that our regulatory model can be inferred for robot control successfully.     

An investigation of neural networks for linear time-series forecasting

This study examines the capability of neural networks for linear time-series forecasting. Using both simulated and real data, the effects of neural network factors such as the number of input nodes and the number of hidden nodes as well as the training sample size are investigated. Results show that neural networks are quite competent in modeling and forecasting linear time series in a variety of situations and simple neural network structures are often effective in modeling and forecasting linear time series.Scope and purposeNeural network capability for nonlinear modeling and forecasting has been established in the literature both theoretically and empirically. The purpose of this paper is to investigate the effectiveness of neural networks for linear time-series analysis and forecasting. Several research studies on neural network capability for linear problems in regression and classification have yielded mixed findings. This study aims to provide further evidence on the effectiveness of neural network with regard to linear time-series forecasting. The significance of the study is that it is often difficult in reality to determine whether the underlying data generating process is linear or nonlinear. If neural networks can compete with traditional forecasting models for linear data with noise, they can be used in even broader situations for forecasting researchers and practitioners.     

LSTM recurrent networks learn simple context-free andcontext-sensitive languages

Previous work on learning regular languages from exemplary training sequences showed that long short-term memory (LSTM) outperforms traditional recurrent neural networks (RNNs). We demonstrate LSTMs superior performance on context-free language benchmarks for RNNs, and show that it works even better than previous hardwired or highly specialized architectures. To the best of our knowledge, LSTM variants are also the first RNNs to learn a simple context-sensitive language, namely a(n)b(n)c(n).     

Evolving Recurrent Neural Network using Cartesian Genetic Programming to Predict The Trend in Foreign Currency Exchange Rates

Forecasting the foreign exchange rate is an uphill task. Numerous methods have been used over the years to develop an efficient and reliable network for forecasting the foreign exchange rate. This study utilizes recurrent neural networks (RNNs) for forecasting the foreign currency exchange rates. Cartesian genetic programming (CGP) is used for evolving the artificial neural network (ANN) to produce the prediction model. RNNs that are evolved through CGP have shown great promise in time series forecasting. The proposed approach utilizes the trends present in the historical data for its training purpose. Thirteen different currencies along with the trade-weighted index (TWI) and special drawing rights (SDR) is used for the performance analysis of recurrent Cartesian genetic programming-based artificial neural networks (RCGPANN) in comparison with various other prediction models proposed to date. The experimental results show that RCGPANN is not only capable of obtaining an accurate but also a computationally efficient prediction model for the foreign currency exchange rates. The results demonstrated a prediction accuracy of 98.872 percent (using 6 neurons only) for a single-day prediction in advance and, on average, 92% for predicting a 1000 days’ exchange rate in advance based on ten days of data history. The results prove RCGPANN to be the ultimate choice for any time series data prediction, and its capabilities can be explored in a range of other fields.     

Nonmonotone Levenberg�CMarquardt training of recurrent neural architectures for processing symbolic sequences

In this paper, we present nonmonotone variants of the Levenberg–Marquardt (LM) method for training recurrent neural networks (RNNs). These methods inherit the benefits of previously developed LM with momentum algorithms and are equipped with nonmonotone criteria, allowing temporal increase in training errors, and an adaptive scheme for tuning the size of the nonmonotone slide window. The proposed algorithms are applied to training RNNs of various sizes and architectures in symbolic sequence-processing problems. Experiments show that the proposed nonmonotone learning algorithms train more effectively RNNs for sequence processing than the original monotone methods.     

Training recurrent networks by Evolino

In recent years, gradient-based LSTM recurrent neural networks (RNNs) solved many previously RNN-unlearnable tasks. Sometimes, however, gradient information is of little use for training RNNs, due to numerous local minima. For such cases, we present a novel method: EVOlution of systems with LINear Outputs (Evolino). Evolino evolves weights to the nonlinear, hidden nodes of RNNs while computing optimal linear mappings from hidden state to output, using methods such as pseudo-inverse-based linear regression. If we instead use quadratic programming to maximize the margin, we obtain the first evolutionary recurrent support vector machines. We show that Evolino-based LSTM can solve tasks that Echo State nets (Jaeger, 2004a) cannot and achieves higher accuracy in certain continuous function generation tasks than conventional gradient descent RNNs, including gradient-based LSTM.     

Recurrent neural network technique for behavioral modeling of power amplifier with memory effects

A new technique for behavioral modeling of power amplifier (PA) with short‐ and long‐term memory effects is presented here using recurrent neural networks (RNNs). RNN can be trained directly with only the input–output data without having to know the internal details of the circuit. The trained models can reflect the behavior of nonlinear circuits. In our proposed technique, we extract slow‐changing signals from the inputs and outputs of the PA and use these signals as extra inputs of RNN model to effectively represent long‐term memory effects. The methodology using the proposed RNN for modeling short‐term and long‐term memory effects is discussed. Examples of behavioral modeling of PAs with short‐ and long‐term memory using both the existing dynamic neural networks and the proposed RNNs techniques are shown. © 2014 Wiley Periodicals, Inc. Int J RF and Microwave CAE 25:289–298, 2015.     

A Neural-Network-Based Model for the Dynamic Simulation of the Tire/Suspension System While Traversing Road Irregularities

This paper deals with the simulation of the tire/suspension dynamics by using recurrent neural networks (RNNs). RNNs are derived from the multilayer feedforward neural networks, by adding feedback connections between output and input layers. The optimal network architecture derives from a parametric analysis based on the optimal tradeoff between network accuracy and size. The neural network can be trained with experimental data obtained in the laboratory from simulated road profiles (cleats). The results obtained from the neural network demonstrate good agreement with the experimental results over a wide range of operation conditions. The NN model can be effectively applied as a part of vehicle system model to accurately predict elastic bushings and tire dynamics behavior. Although the neural network model, as a black-box model, does not provide a good insight of the physical behavior of the tire/suspension system, it is a useful tool for assessing vehicle ride and noise, vibration, harshness (NVH) performance due to its good computational efficiency and accuracy.      

Applicability of feed-forward and recurrent neural networks to Boolean function complexity modeling

In this paper, we present the feed-forward neural network (FFNN) and recurrent neural network (RNN) models for predicting Boolean function complexity (BFC). In order to acquire the training data for the neural networks (NNs), we conducted experiments for a large number of randomly generated single output Boolean functions (BFs) and derived the simulated graphs for number of min-terms against the BFC for different number of variables. For NN model (NNM) development, we looked at three data transformation techniques for pre-processing the NN-training and validation data. The trained NNMs are used for complexity estimation for the Boolean logic expressions with a given number of variables and sum of products (SOP) terms. Both FFNNs and RNNs were evaluated against the ISCAS benchmark results. Our FFNNs and RNNs were able to predict the BFC with correlations of 0.811 and 0.629 with the benchmark results, respectively.     

基于周期性建模的时间序列预测方法及电价预测研究

时间序列数据广泛存在于人类的生产生活中, 通常具有复杂的非线性动态和一定的周期性. 与传统的时间序列分析方法相比, 基于深度学习的方法更能捕捉数据的深层特性, 对具有复杂非线性的时间序列有较好的建模效果. 为了在神经网络中显式地建模时间序列数据的周期性和趋势性, 本文在循环神经网络的基础上引入了周期损失和趋势损失, 建立了基于周期性建模和多任务学习的时间序列预测模型. 将模型应用到欧洲能源交易所法国市场的能源市场价格预测中, 结果表明周期损失和趋势损失能够提高神经网络的泛化能力, 并提高预测时间序列趋势的精度.     

Stability analysis of discrete-time recurrent neural networks based on standard neural network models

In order to conveniently analyze the stability of various discrete-time recurrent neural networks (RNNs), including bidirectional associative memory, Hopfield, cellular neural network, Cohen-Grossberg neural network, and recurrent multiplayer perceptrons, etc., the novel neural network model, named standard neural network model (SNNM) is advanced to describe this class of discrete-time RNNs. The SNNM is the interconnection of a linear dynamic system and a bounded static nonlinear operator. By combining Lyapunov functional with S-Procedure, some useful criteria of global asymptotic stability for the discrete-time SNNMs are derived, whose conditions are formulated as linear matrix inequalities. Most delayed (or non-delayed) RNNs can be transformed into the SNNMs to be stability analyzed in a unified way. Some application examples of the SNNMs to the stability analysis of the discrete-time RNNs shows that the SNNMs make the stability conditions of the RNNs easily verified.     

A System for Robotic Heart Surgery that Learns to Tie Knots Using Recurrent Neural Networks

Tying suture knots is a time-consuming task performed frequently during minimally invasive surgery (MIS). Automating this task could greatly reduce total surgery time for patients. Current solutions to this problem replay manually programmed trajectories, but a more general and robust approach is to use supervised machine learning to smooth surgeon-given training trajectories and generalize from them. Since knot tying generally requires a controller with internal memory to distinguish between identical inputs that require different actions at different points along a trajectory, it would be impossible to teach the system using traditional feedforward neural nets or support vector machines. Instead we exploit more powerful, recurrent neural networks (RNNs) with adaptive internal states. Results obtained using long short-term memory RNNs trained by the recent Evolino algorithm show that this approach can significantly increase the efficiency of suture knot tying in MIS over preprogrammed control.     

Combining flat and structured representations for fingerprint classification with recursive neural networks and support vector machines

We present new fingerprint classification algorithms based on two machine learning approaches: support vector machines (SVMs) and recursive neural networks (RNNs). RNNs are trained on a structured representation of the fingerprint image. They are also used to extract a set of distributed features of the fingerprint which can be integrated in the SVM. SVMs are combined with a new error-correcting code scheme. This approach has two main advantages: (a) It can tolerate the presence of ambiguous fingerprint images in the training set and (b) it can effectively identify the most difficult fingerprint images in the test set. By rejecting these images the accuracy of the system improves significantly. We report experiments on the fingerprint database NIST-4. Our best classification accuracy is of 95.6 percent at 20 percent rejection rate and is obtained by training SVMs on both FingerCode and RNN-extracted features. This result indicates the benefit of integrating global and structured representations and suggests that SVMs are a promising approach for fingerprint classification.     

Combining empirical mode decomposition and deep recurrent neural networks for predictive maintenance of lithium-ion battery

Predictive maintenance of lithium-ion batteries has been one of the popular research subjects in recent years. Lithium-ion batteries can be used as the energy supply for industrial equipment, such as automated guided vehicles and battery electric vehicles. Predictive maintenance plays an important role in the application of smart manufacturing. This mechanism can provide different levels of pre-diagnosis for machines or components. Remaining useful life (RUL) prediction is crucial for the implementation of predictive maintenance strategies. RUL refers to the estimated useful life remaining before the machine cannot operate after a certain period of operation. This study develops a hybrid data science model based on empirical mode decomposition (EMD), grey relational analysis (GRA), and deep recurrent neural networks (RNN) for the RUL prediction of lithium-ion batteries. The EMD and GRA methods are first adopted to extract the characteristics of time series data. Then, various deep RNNs, including vanilla RNN, gated recurrent unit, long short-term memory network (LSTM), and bidirectional LSTM, are established to forecast state of health (SOH) and the RUL of lithium-ion batteries. Bayesian optimization is also used to find the best hyperparameters of deep RNNs. Experimental results with the lithium-ion batteries data of NASA Ames Prognostics Data Repository show that the proposed hybrid data science model can accurately predict the SOH and RUL of lithium-ion batteries. The LSTM network has the optimal results. The proposed hybrid data science model with multiple artificial intelligence-based technologies also demonstrates critical digital-technology enablers for digital transformation of smart manufacturing and transportation.