Improved Stability Analysis for Delayed Neural Networks

Trainable Delays in Time Delay Neural Networks for Learning Delayed Dynamics

In this article, the connection between time delay systems and time delay neural networks (TDNNs) is presented from a continuous-time perspective. TDNNs are utilized to learn the nonlinear dynamics of time delay systems from trajectory data. The concept of TDNN with trainable delay (TrTDNN) is established, and training algorithms are constructed for learning the time delays and the nonlinearities simultaneously. The proposed techniques are tested on learning the dynamics of autonomous systems from simulation data and on learning the delayed longitudinal dynamics of a connected automated vehicle (CAV) from real experimental data.

Adaptive event-triggered synchronization of neural networks under stochastic cyber-attacks with application to Chua's circuit

This paper focuses on the synchronization control problem for neural networks (NNs) subject to stochastic cyber-attacks. Firstly, an adaptive event-triggered scheme (AETS) is adopted to improve the utilization rate of network resources, and an output feedback controller is constructed for improving the performance of the system subject to the conventional deception attack and accumulated dynamic cyber-attack. Secondly, the synchronization problem of master-slave NNs is transformed into the stability analysis problem of the synchronization error system. Thirdly, by constructing a customized Lyapunov-Krasovskii functional (LKF), the adaptive event-triggered output feedback controller is designed to ensure the synchronization error system is asymptotically stable with a given H∞ performance index. Lastly, in the simulation part, two examples, including Chua's circuit, illustrate the feasibility and universality of the related technologies in this paper.

Novel Extended State Observer Design for Uncertain Nonlinear Systems via Refined Dynamic Event-Triggered Communication Protocol

In this article, an extended state observer (ESO) design problem is investigated for uncertain nonlinear systems subject to limited network bandwidth. First, for rational information exchange scheduling, a dynamic event-triggered (DET) communication protocol is proposed. Different from the traditional static event-triggered strategies with fixed thresholds, an internal dynamic variable is introduced to be adaptively adjusted by a dual-directional regulating mechanism. Thus, more desirable tradeoff between observation performance and communication resource efficiency is achieved. Second, inspired by our early work on Takagi-Sugeno fuzzy ESO (TSFESO), a novel paradigm of event-triggered TSFESO is initially proposed. Third, under the DET mechanism, the TSFESO design approach is derived to carry out exponential convergence for estimation error dynamics. Finally, the effectiveness of the proposed method is verified by numerical examples. The nonlinear estimating efficiency and linear numerical tractability are integrated in TSFESO. In addition, a generalized ESO formulation is developed to allow some nonadditive uncertainties incompatible with total disturbance, such as improved event-triggered strategy, and thus, the application sphere of ESO is further expanded.

Wavelet Probabilistic Neural Networks

In this article, a novel wavelet probabilistic neural network (WPNN), which is a generative-learning wavelet neural network that relies on the wavelet-based estimation of class probability densities, is proposed. In this new neural network approach, the number of basis functions employed is independent of the number of data inputs, and in that sense, it overcomes the well-known drawback of traditional probabilistic neural networks (PNNs). Since the parameters of the proposed network are updated at a low and constant computational cost, it is particularly aimed at data stream classification and anomaly detection in off-line settings and online environments where the length of data is assumed to be unconstrained. Both synthetic and real-world datasets are used to assess the proposed WPNN. Significant performance enhancements are attained compared to state-of-the-art algorithms.

H∞ State Estimation for Neural Networks With General Activation Function and Mixed Time-Varying Delays

This article deals with H_∞ state estimation of neural networks with mixed delays. In order to make full use of delay information, novel delay-product Lyapunov-Krasovskii functional (LKF) by using parameterized delay interval is first constructed. Then, generalized free-weighting-matrix integral inequality is used to estimate the derivative of LKF to reduce the conservatism. Also, a more general activation function is further applied by combining with parameterized delay interval in order to obtain a more accurate estimator model. Finally, sufficient conditions are derived to confirm that the estimation error system is asymptotically stable with a prescribed H_∞ performance. Numerical examples are simulated to show the benefits of our proposed method.

Bilinear neural network with 3-D attention for brain decoding of motor imagery movements from the human EEG

Deep learning has achieved great success in areas such as computer vision and natural language processing. In the past, some work used convolutional networks to process EEG signals and reached or exceeded traditional machine learning methods. We propose a novel network structure and call it QNet. It contains a newly designed attention module: 3D-AM, which is used to learn the attention weights of EEG channels, time points, and feature maps. It provides a way to automatically learn the electrode and time selection. QNet uses a dual branch structure to fuse bilinear vectors for classification. It performs four, three, and two classes on the EEG Motor Movement/Imagery Dataset. The average cross-validation accuracy of 65.82%, 74.75%, and 82.88% was obtained, which are 7.24%, 4.93%, and 2.45% outperforms than the state-of-the-art, respectively. The article also visualizes the attention weights learned by QNet and shows its possible application for electrode channel selection.

Adaptive Synchronization of Reaction-Diffusion Neural Networks and Its Application to Secure Communication

This paper is mainly concerned with the synchronization problem of reaction-diffusion neural networks (RDNNs) with delays and its direct application in image secure communications. An adaptive control is designed without a sign function in which the controller gain matrix is a function of time. The synchronization criteria are established for an error model derived from master-slave models through solving the set of linear matrix inequalities derived by constructing the suitable novel Lyapunov-Krasovskii functional candidate, Green's formula, and Wirtinger's inequality. If the proposed sufficient conditions are satisfied, then the global asymptotic synchronization of the error model is guaranteed. The numerical illustrations are provided to demonstrate the validity of the derived synchronization criteria. In addition, the role of system parameters is picturized through the chaotic nature of RDNNs and those unprecedented solutions is utilized to promote better security of image transactions. As is evident, the enhancement of image encryption algorithm is designed with two levels, namely, image watermarking and diffusion process. The contributions of this paper are discussed as concluding remarks.

Stability Analysis for Delayed Neural Networks via Improved Auxiliary Polynomial-Based Functions

This brief is concerned with stability analysis for delayed neural networks (DNNs). By establishing polynomials and introducing slack variables reasonably, some improved delay-product type of auxiliary polynomial-based functions (APFs) is developed to exploit additional degrees of freedom and more information on extra states. Then, by constructing Lyapunov-Krasovskii functional using APFs and integrals of quadratic forms with high order scalar functions, a novel stability criterion is derived for DNNs, in which the benefits of the improved inequalities are fully integrated and the information on delay and its derivative is well reflected. By virtue of the advantages of APFs, more desirable performance is achieved through the proposed approach, which is demonstrated by the numerical examples.