Neural Networks最新文献

This paper presents cutting-edge advancements in exponential synchronization and encryption techniques, focusing on Quaternion-Valued Artificial Neural Networks (QVANNs) that incorporate two-sided coefficients. The study introduces a novel approach that harnesses the Cayley-Dickson representation method to simplify the complex equations inherent in QVANNs, thereby enhancing computational efficiency by exploiting complex number properties. The study employs the Lyapunov theorem to craft a resilient control system, showcasing its exponential synchronization by skillfully regulating the Lyapunov function and its derivatives. This management ensures the stability and synchronization of the network, which is crucial for reliable performance in various applications. Extensive numerical simulations are conducted to substantiate the theoretical framework, providing empirical evidence supporting the presented design and proofs. Furthermore, the paper explores the practical application of QVANNs in the encryption and decryption of color images, showcasing the network's capability to handle complex data processing tasks efficiently. The findings of this research not only contribute significantly to the development of complex artificial neural networks but pave the way for further exploration into systems with diverse delay types.

Memristors are commonly used as the connecting parts of neurons in brain-like neural networks. The memristors, unlike the existing literature, possess the capability to function as both self-connected synaptic weights and interconnected synaptic weights, thereby enabling the generation of intricate initials-regulated plane coexistence behaviors. To demonstrate this dynamical effect, a Hopfield neural network with two-memristor-interconnected neurons (TMIN-HNN) is proposed. On this basis, the stability distribution of the equilibrium points is analyzed, the related bifurcation behaviors are studied by utilizing some numerical simulation methods, and the plane coexistence behaviors are proved theoretically and revealed numerically. The results clarify that TMIN-HNN not only exhibits complex bifurcation behaviors, but also has initials-regulated plane coexistence behaviors. In particular, the coexistence attractors can be switched to different plane locations by the initial states of the two memristors. Finally, a digital experiment device is developed based on STM32 hardware board to verify the initials-regulated plane coexistence attractors.

This article investigates the problem of adaptive fixed-time optimal consensus tracking control for nonlinear multiagent systems (MASs) affected by actuator faults and input saturation. To achieve optimal control, reinforcement learning (RL) algorithm which is implemented based on neural network (NN) is employed. Under the actor-critic structure, an innovative simple positive definite function is constructed to obtain the upper bound of the estimation error of the actor-critic NN updating law, which is crucial for analyzing fixed-time stabilization. Furthermore, auxiliary functions and estimation laws are designed to eliminate the coupling effects resulting from actuator faults and input saturation. Meanwhile, a novel event-triggered mechanism (ETM) that incorporates the consensus tracking errors into the threshold is proposed, thereby effectively conserving communication resources. Based on this, a fixed-time event-triggered control scheme grounded in RL is proposed through the integration of the backstepping technique and fixed-time theory. It is demonstrated that the consensus tracking errors converge to a specified range in a fixed time and all signals within the closed-loop systems are bounded. Finally, simulation results are provided to verify the effectiveness of the proposed control strategy.

The widespread development and use of neural networks have significantly enriched a wide range of computer algorithms and promise higher speed at lower cost. However, the imitation of neural networks by means of modern computing substrates is highly inefficient, whereas physical realization of large scale networks remains challenging. Fortunately, delayed-feedback oscillators, being much easier to realize experimentally, represent promising candidates for the empirical implementation of neural networks and next generation computing architectures. In the current research, we demonstrate that coupled bistable delayed-feedback oscillators emulate a multilayer network, where one single-layer network is connected to another single-layer network through coupling between replica nodes, i.e. the multiplex network. We show that all the aspects of the multiplexing impact on wavefront propagation and stochastic resonance identified in multilayer networks of bistable oscillators are entirely reproduced in the dynamics of time-delay oscillators. In particular, varying the coupling strength allows suppressing and enhancing the effect of stochastic resonance, as well as controlling the speed and direction of both deterministic and stochastic wavefront propagation. All the considered effects are studied in numerical simulations and confirmed in physical experiments, showing an excellent correspondence and disclosing thereby the robustness of the observed phenomena.

According to the developmental process of infants, cognitive abilities are divided into four stages: the Exploration Stage (ES), the Mapping Stage (MS), the Phenomena-causality Stage (PCS), and the Essence-causality Stage (ECS). The MS is a training of the consecutive characteristics of events, similar to a deep learning model; the PCS is a process that symbolizes the input and output of the mapping training, and uses these symbols as the input or output of the mapping training again. After training, the next possible symbol can be predicted, which is equivalent to recognizing the essence. Expressing the essence itself with a function in the ECS represents entering the scope of science. To illustrate the above process, take the evolution journey of an insectoid with only visual and compositional detection capabilities as an example. Without the need for additional learning algorithm programming, the insectoid evolves according to the Cognitive Process and Information Processing Model and can develop its own independent symbol system. The ability to develop its own unique symbolic system actually indicates the birth of an agent.

Traffic flow prediction is the foundation of intelligent traffic management systems. Current methods prioritize the development of intricate models to capture spatio-temporal correlations, yet they often neglect the exploitation of latent features within traffic flow. Firstly, the correlation among different road nodes exhibits dynamism rather than remaining static. Secondly, traffic data exhibits evident periodicity, yet current research lacks the exploration and utilization of periodic features. Lastly, current models typically rely solely on historical data for modeling, resulting in the limitation of accurately capturing future trend changes in traffic flow. To address these findings, this paper proposes a Periodic Dynamic Graph to Sequence Model (PDG2Seq) for traffic flow prediction. PDG2Seq consists of the Periodic Feature Selection Module (PFSM) and the Periodic Dynamic Graph Convolutional Gated Recurrent Unit (PDCGRU) to further extract the spatio-temporal features of the dynamic real-time traffic. The PFSM extracts learned periodic features using time points as indices, while the PDCGRU leverages the extracted periodic features from the PFSM and dynamic features from traffic flow to generate a Periodic Dynamic Graph for extracting spatio-temporal features. In the decoding phase, PDG2Seq utilizes periodic features corresponding to the prediction target to capture future trend changes, leading to more accurate predictions. Comprehensive experiments conducted on four large-scale datasets substantiate the superiority of PDG2Seq over existing state-of-the-art baselines. Related codes are available at https://github.com/wengwenchao123/PDG2Seq.

Recently, heterogeneous graphs have attracted widespread attention as a powerful and practical superclass of traditional homogeneous graphs, which reflect the multi-type node entities and edge relations in the real world. Most existing methods adopt meta-path construction as the mainstream to learn long-range heterogeneous semantic messages between nodes. However, such schema constructs the node-wise correlation by connecting nodes via pre-computed fixed paths, which neglects the diversities of meta-paths on the path type and path range. In this paper, we propose a meta-path-based semantic embedding schema, which is called Heterogeneous Graph Embedding with Dual Edge Differentiation (HGE-DED) to adequately construct flexible meta-path combinations thus learning the rich and discriminative semantic of target nodes. Concretely, HGE-DED devises a Multi-Type and multi-Range Meta-Path Construction (MTR-MP Construction), which covers the comprehensive exploration of meta-path combinations from path type and path range, expressing the diversity of edges at more fine-grained scales. Moreover, HGE-DED designs the semantics and meta-path joint guidance, constructing a hierarchical short- and long-range relation adjustment, which constrains the path learning as well as minimizes the impact of edge heterophily on heterogeneous graphs. Experimental results on four benchmark datasets demonstrate the effectiveness of HGE-DED compared with state-of-the-art methods.

The detection of infrared small targets is getting more and more attention, and has a wider application in both military and civilian fields. The traditional infrared small target detection methods heavily rely on the setting of manual features, and the deep learning-based method easily lose the targets in deep layers due to several downsampling operations. To handle this problem, we design multi-level multi-patch multi-receptive multi-dimensional attention network (M4Net) to achieve information interaction among high-level and low-level features for maintaining target contour and location detail. Multi-level feature extraction module (MFEM) with multilayer vision transformer (ViT) is introduced under the encoder-decoder framework to fuse multi-scale features. Multi-patch attention module (MPAM) and multi-receptive field module (MRFM) are proposed to capture and enhance the feature information. Multi-dimension interactive module (MDIM) is designed to connect the attention mechanism on multiscale features to enhance the network's leaning ability. Finally, the extensive experiments carried out on infrared small target detection dataset demonstrate that our method achieves better performance compared to other methods.

Multivariate time series classification is of great importance in practical applications and is a challenging task. However, deep neural network models such as Transformers exhibit high accuracy in multivariate time series classification but lack interpretability and fail to provide insights into the decision-making process. On the other hand, traditional approaches based on decision tree classifiers offer clear decision processes but relatively lower accuracy. Swin Transformer (ST) addresses these issues by leveraging self-attention mechanisms to capture both fine-grained local patterns and global patterns. It can also model multi-scale feature representation learning, thereby providing a more comprehensive representation of time series features. To tackle the aforementioned challenges, we propose ST-Tree with interpretability for multivariate time series classification. Specifically, the ST-Tree model combines ST as the backbone network with an additional neural tree model. This integration allows us to fully leverage the advantages of ST in learning time series context while providing interpretable decision processes through the neural tree. This enables researchers to gain clear insights into the model's decision-making process and extract meaningful interpretations. Through experimental evaluations on 10 UEA datasets, we demonstrate that the ST-Tree model improves accuracy in multivariate time series classification tasks and provides interpretability through visualizing the decision-making process across different datasets.

Few-shot image generation aims at generating novel images for the unseen category when given K images from the same category. Despite significant advancements in existing few-shot image generation methods, great challenges remain regarding the quality and diversity of the generated images. This issue stems from the model's struggle to fully comprehend the semantic content of images and extract sufficiently semantic representations. To address these issues, we propose a semantic mask reconstruction (SMR) and category semantic learning (CSL) method for few-shot image generation. Specifically, SMR performs mask reconstruction in a high-level semantic space and designs a strategy for dynamically adjusting the mask ratio, which increases the difficulty of the generation tasks by gradually increasing the mask ratio to enhance the learning ability of the discriminator, thereby prompting the generator to learn more critical features relevant to the generation task. In addition, CSL introduces a triplet loss to optimize the distance between the generated image, its corresponding input image, and input images of other categories. This encourages the generative model to discern subtle differences between categories, thereby achieving more fine-grained generation and improving the fidelity of generated images. Both SMR and CSL can function as plug-and-play modules. Extensive experimental results across three standard datasets demonstrate that the SMR-CSL outperforms other methods in terms of the quality and diversity of the generated images. Furthermore, the results of downstream classification experiments verify that the images generated by the proposed method can effectively assist downstream classification tasks.