Compared to image frames, the event stream captured by event cameras possesses the advantages of being high temporal resolution, sparsity and asynchronism. However, these characteristics also make it highly susceptible to noise which negatively affects the performance of downstream tasks. Existing event denoising methods are either too straightforward for varying noise-ratio scenes, or too complex, resulting in low efficiency and difficult to use in practice. In search of a denoising method that combines both performance and efficiency, we propose a lightweight and real-time event denoising algorithm based on the Spiking Neural Networks (SNN). Specifically, we introduce the Threshold-Limited PLIF neuron model, which leverages membrane potentials to capture the spatio-temporal correlations essential for effective denoising. With this neuron as the fundamental component, we design a frame-by-frame denoising network, called DeSNN. The proposed architecture utilizes the SNN architecture to integrate the spatio-temporal information from input event frames, and subsequently generates denoised outputs. Our method allows the entire processing pipeline to be maintained in the spiking data format, thereby fully exploiting the strengths of both dynamic vision sensors (DVS) and SNNs. Extensive experimental results demonstrated that our method achieves both high performance and high efficiency, which is beneficial for real-time downstream tasks. Furthermore, we have implemented our method in several simple real-world scenarios, illustrating its practical applicability and potential for deployment.

Medical image translation plays a crucial role in assisting clinical diagnosis by enabling cross-modal synthesis (e.g., computed tomography to magnetic resonance imaging) and super-resolution, effectively addressing clinical challenges, such as radiation exposure, prolonged scan times, and allergic reactions to contrast agents. However, existing approaches primarily focus on developing specialized models for specific tasks, limiting their adaptability across different applications. Developing a general model capable of handling arbitrary medical image translation tasks not only enhances cross-domain generalization but also aligns with the broader trend of artificial intelligence evolving from specialized to general-purpose solutions. Achieving such one-for-all model, however, presents three key challenges: first, varying task complexity; second, modality discrepancies; third, structural variations across anatomical regions within the same modality. To tackle the first challenge, we utilize an advanced diffusion-based training paradigm to endow the denoising model with extensive pattern coverage capabilities, thereby handling tasks of varying difficulty levels. Subsequently, a general diffusion Transformer incorporating a fuzzy mixture-of-experts (FMoE) module and an entropy-guided attention soft prompt (EASP) module is proposed. The FMoE module, equipped with nonlinear modeling capabilities, is designed to address modality discrepancies, while the EASP module is employed to enhance the model's perception of structural variations in images. Extensive qualitative and quantitative experiments demonstrate the effectiveness of the proposed model in the arbitrary medical image translation task.

To address fragmented multimodal perception, high energy consumption in hardware deployment, and delayed responses in anomalous driving scenarios, this work proposes a coupled gradient-evolutionary learning framework deployed on sparse memristive neuromorphic networks for robust edge intelligence. The framework fuses physiological, visual, and auditory modalities and is validated across multimodal benchmarks on image, audio, and EEG tasks for real-time driving-safety monitoring, forming a closed-loop perception-analysis-decision pipeline that improves the reliability of safety-critical decisions. At the hardware level, neuromorphic processing units are constructed using two-dimensional material-based memristors. Leveraging in-memory computing and parallel processing capabilities of the memristive architecture, the proposed framework achieves energy-efficient classification of multimodal signals. At the algorithm level, a cross-species-inspired gradient-evolutionary architecture integrates local visual-cortex-inspired CNN learning for traffic-scene parsing with global Darwinian population evolution. Memristor write noise is utilized as a functional perturbation to drive the evolutionary training, which improves classification accuracy under quantization noise and enhances robustness against hardware non-idealities. With the synergy between evolutionary training and memristor-aware low-bit quantization, the framework exhibits enhanced natural sparsity and achieves tens-of-milliseconds inference latency, tens of milliwatts power consumption, and an energy cost of hundreds of microjoules per inference, resulting in about 5.8 ×  energy savings compared to conventional von Neumann edge computing architectures. Overall, this work provides a low-carbon, robust, and scalable edge intelligence solution for road safety decision-making in human-vehicle-environment systems, demonstrating the potential of neuromorphic computing for supporting carbon neutrality in transportation.

This paper presents a Sine-Lucas Oscillation Particle Swarm Optimization (SLOPSO) XGBoost algorithm for landslide susceptibility prediction. SLOPSO extends canonical PSO by introducing an oscillation factor constructed from a sine function and the Lucas number sequence into the position update, which raises swarm diversity and improves the global search. SLOPSO is then applied to tune six XGBoost hyperparameters using K-fold cross-validation accuracy as the fitness function. In experimental evaluation, SLOPSO-XGBoost reaches a mean test Accuracy of 0.8113, F1 of 0.8182, and AUC of 0.8856 over 10 independent runs, outperforming standard PSO-XGBoost and the default Random Forest, SVM, and XGBoost baselines on every metric. The experimental results demonstrate that SLOPSO is capable of automatically tuning the hyperparameters of XGBoost, and that SLOPSO-XGBoost able to provide high precision solution for landslide event prediction.

Event cameras are innovative neuromorphic sensors that capture dynamic changes in scenes, recording millions of events per second. Recent sparse computational models on event stream recognition have achieved notable successes by utilizing graph convolution or attention mechanisms to model local dependencies. However, these methods face significant challenges in constructing effective event representations and modeling global dependencies when confronted with millions of events in spacetime. To surmount the above challenges, we present a novel voxel-based State Space Model (SSM) network, termed VeMamba, which can establish multi-scale spatiotemporal dependencies in event voxels with linear complexity. Specifically, we design a time-aware enhanced voxelization method that enriches the spatiotemporal expression within event voxels while preserving sparse computation. Then, we propose a multi-scale modeling module with linear complexity that integrates local attention into the global dual-scale SSMs to establish spatiotemporal dependencies from local to global within serialized voxels. Furthermore, leveraging a hierarchical structure grounded in voxel merging, we can extract deep semantic and motion cues from the voxels. Extensive experiments demonstrate that VeMamba achieves state-of-the-art (SOTA) performance with low model complexity and computational cost on event stream recognition tasks.

Dear Editor, This letter concerns the design of sliding mode control (SMC) for semi-Markov switching systems with time-varying transmission and impulse delay. The difficulties of this problem are: 1) Time-varying transmission and impulse delay bring more nonlinear dynamic characteristics and lag effects; 2) Semi-Markov mode switching introduces uncertainty; 3) The reachable stage and sliding stage are affected by two types of impulses in the system, which increases the complexity of theoretical derivation.

Spatial auditory localization enables barn owls to accurately locate and capture prey in low-light conditions. However, most existing works simulating this ability fail to integrate horizontal and vertical auditory localization, rarely combine it with navigation tasks, and rely mostly on non-neuromorphic computing methods. To address these limitations, this work proposes a memristive hybrid neural network (HyNN) circuit for dual-angle auditory localization and online learning-based navigation. The circuit comprises four modules. The horizontal and vertical angle localizing modules respectively sparsely encode the interaural time difference and interaural level difference into the firing intensity of neurons that represent spatial information. The winner flag generation and learning rate control modules selectively and adaptively adjust weight updates. For neuromorphic implementation, the HyNN is fully deployed on a physical memristive computing-in-memory (CIM) platform, achieving approximately 8.77× lower energy consumption than a novel low-power FPGA accelerator. This work offers a novel and energy-efficient method for autonomous edge agent to perform auditory localization-based navigation in visually limited spaces.

Memristors are regarded as excellent carriers for simulating biological synapses due to their unique memory properties. When employing memristive neuron models to simulate biological neuronal firing behaviors, the complexity of firing activities is a critical measure in assessing model performance. However, the diversity and complexity of firing patterns generated by individual neurons or memristor-coupled single neurons typically suffer from certain limitations. To solve this problem, this paper employs a universal discrete memristor for synaptic coupling between two Rulkov neurons to build a memristive coupled dual Rulkov neuron model (MCDRNM). By switching the memristor’s internal parameters, the MCDRNM can generate a controllable number of multi-type hyperchaotic firing patterns, including multi-foot-shaped and multi-pigeon-shaped patterns. Through adjustments of the initial conditions of the memristor, the MCDRNM not only exhibits heterogeneous firing multistability but also induces homogeneous firing multistability with infinitely coexisting single-foot-shaped/pigeon-shaped hyperchaotic firing patterns. Furthermore, firing pattern transitions are also revealed in this MCDRNM. Meanwhile, STM32 hardware experiments are conducted, demonstrating the physical realizability of the model. Finally, based on the hyperchaotic firing sequences of the MCDRNM, an efficient image encryption method suitable for the Internet of Medical Things (IoMT) is designed. The evaluation results demonstrate that this encryption scheme is as efficient as it is secure. It achieves high encryption efficiency while maintaining strong resistance to attacks.

Medical institutions store a vast amount of patient information, and various types of medical data face severe security challenges in cloud storage environments. This paper proposes an efficient multi-type medical multimedia data encryption scheme based on the memristive Hopfield neural network (MHNN). First, a class of MHNNs is constructed, which exhibits isomorphic extreme multistability under different initial conditions. The systems are capable of generating large-scale coexisting chaotic attractors, whose spatial positions, vortex numbers, and amplitudes can be independently regulated, thereby significantly expanding the generation capacity and diversity of chaotic sequences. Based on this, a multi-type data encryption algorithm is designed, which uniformly encodes and integrates different types of medical data into a secure transmission structure. Furthermore, by exploiting multistable characteristics of the MHNN, a two-layer key system composed of a master key and a selection key is established, enabling a secure medical data scheme that supports multi-party collaborative annotation. Experimental results demonstrate that the proposed encryption scheme achieves excellent performance in both security and efficiency, providing a novel and effective solution for secure storage and collaborative processing of multi-type medical multimedia data.

Event cameras have garnered widespread attention in motion recognition and detection due to their advantages of low latency and fewer privacy concerns. To adapt existing learning models for event data, current leading works either accumulate event streams into dense frames or voxel structures and process them by standard convolutional neural network (CNN) or vision transformer (ViT) architectures, or employ lightweight point-based networks to learn sparse representations. However, the former inevitably sacrifices event sparsity and spatial proximity, increasing computational burden, while the latter typically relies on a local feature extractor, which limits scalability and long-range feature interaction. To address these issues, we propose AGFI-Net, a novel framework leveraging adaptive graph construction and multi-serialization patterns. Initially, we transform event streams into event graphs by performing adaptive downsampling and denoising to preserve the sparsity of the raw data. Then, we utilize space-filling curves (SFCs) to map unordered 3D vertices into 1D sequences, enabling structured feature learning while maintaining the benefits of spatiotemporal locality. To enhance vertex encoding, we introduce curve shuffle attention (CSA) blocks with linear positional embeddings during the feature interaction phase, effectively compensating for the lack of global context. Experimental results show that our approach surpasses the highest levels on action recognition and tiny object detection benchmarks, validating its superior performance. Our source code is available at: https://github.com/hust-fstudy/AGFI-Net.

The secure sharing of multiple medical images in Internet of Things environments faces challenges related to privacy leakage, unauthorized access, and copyright trace-ability. Existing multi-image encryption schemes mainly focus on confidentiality, commonly employ a unified key structure, and design encryption and watermark authentication separately, making it difficult to support patient-specific recovery and non-embedding authentication simultaneously. To address these limitations, this paper proposes a secure medical multi-image transmission framework based on a memristive electromagnetic-radiation Hopfield neural network (MERHNN). The MERHNN generates non-uniform multi-double-scroll attractor clusters, and its multistability is exploited to construct a hierarchical key architecture. The master key controls global protection, whereas patient-specific subkeys generate local key streams. A book-page-inspired three-dimensional cylindrical mapping combines intra-image scrambling, cross-image global scrambling, and bidirectional diffusion to jointly protect the image set. A zero-watermarking mechanism integrating Radon-based orientation estimation, adaptive sampling, spatial–frequency feature fusion, and BCH coding enables authentication without modifying the original images. Experimental results confirm lossless image recovery and demonstrate effective statistical properties, resistance to differential attacks, and robust watermark extraction under the tested attacks. The proposed framework therefore supports differentiated authorization and reliable authentication for sharing multiple medical images.

The parallax error (PE) significantly deteriorates the spatial resolution and imaging quality of positron emission tomography (PET) scanners. Existing PE correction methods either rely on depth decoding detectors in hardware which increases development costs, or optimize the system response matrix (SRM) in software providing limited compensation for PE. This work proposed a novel PE correction method in projection space based on deep learning (DL), consisting of two steps. First, the sinogram affected by PE was processed by a neural network (PEC-Net). The corrected sinogram output from the PEC-Net was then reconstructed to an improved image. To generate ideal PE-corrected labels, we synthesized training data using Monte Carlo (MC) simulation-based SRMs as forward projectors. The proposed method was validated using simulation data and real data. Experimental results show that the proposed method effectively eliminated artifacts caused by PE, and the reconstructed images of simulation data outperformed those obtained at 4 mm depth of interaction (DOI) resolution in terms of structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR). The PEC-Net may provide a low-cost, high-performance, software-based PE correction method for PET scanners without DOI measurement.

Designing and deploying highly secure and efficient encryption algorithms for Internet of Things (IoT) devices, particularly resource-constrained medical devices, poses a significant challenge. Chaotic systems, with their sensitivity to initial conditions and capacity to generate pseudorandom signals, offer a promising solution to address the limitations of image encryption in IoT devices. Based on a novel memristive multi-attractor Hénon map model (MMHM) that is able to generate hyperchaotic signals, this article proposes a lightweight image encryption scheme for medical Internet of Medical Things (IoMT) devices, effectively addressing numerous deficiencies in existing approaches. To overcome the leakage risks present in existing schemes during key distribution, the initial key is encrypted using an asymmetric encryption algorithm. A key update mechanism is also employed, guaranteeing that each encrypted image is assigned a unique key, which effectively nullifies the risk of differential attacks. Additionally, a geometric transformation-based disruption algorithm is introduced, achieving exceptionally high levels of chaos with minimal computational overhead while providing robust resistance to cropping attacks. Notably, the solution is deployed on a digital circuit platform based on the STM32 microcontroller. Experimental results demonstrate that this method significantly outperforms traditional approaches in resisting typical attacks. It effectively compensates for vulnerabilities in the key distribution and key update mechanisms of conventional schemes, substantially enhancing both cryptographic security and efficiency.

Silicone adhesives with their excellent viscoelasticity and aging resistance are often widely used as interconnection materials in fields such as electronic packaging. However, due to the difference in thermal expansion coefficients, the packaging structure may experience adhesive failure caused by thermal stress. In this paper, a temperature-dependent viscoelastic model of silicone adhesives was developed to study the mechanical behavior changes of an adhesive and relevant parameter expressions were provided. Additionally, deep learning and FEA were employed for comparative verification of the model. The results indicate that both the viscoelastic model and the deep learning model used in this study can capture the trend of temperature's influence on the adhesive. Both models successfully capture the temperature-dependent variation law of the adhesive's mechanical properties in both the low-temperature region below the critical temperature Tref and the high-temperature region above the critical temperature Tref. However, the prediction results indicate that the viscoelastic model is more consistent with the variation law of the experimental data. Meanwhile, the finite element simulation results are generally consistent with those of the theoretical model. This model can provide theoretical guidance for studying the mechanical behavior of other adhesives operating at different temperatures.

For group containment control (GCC) problem, the presence of multiple subgroups and multiple leaders in each subgroup poses challenges for topology selection and problem analysis. This paper presents the endeavor to investigate the GCC problem of multi-agent systems with unknown dynamics, focusing specifically on optimizing its performance. First, the group neighborhood containment error is defined, and it is proven that the GCC can be achieved by converging the error to zero under the well-defined communication topology. Meanwhile, a feasible communication topology selection algorithm is proposed for GCC problem. Then, considering the optimization of control performance, the optimal GCC problem is formulated via optimality principle and graphical game by defining the local performance index for each agent. Based on adaptive dynamic programming, two novel online model-free methods are developed for solving optimal GCC problem, together with rigorous mathematical analysis. It is demonstrated that under the two methods, the optimal control policy can be learned using only system operation data, without requiring knowledge of system dynamics, thereby achieving the optimal GCC. Finally, corresponding simulation examples are executed to demonstrate the capacity of the developed methods.

With the advancement of intelligent healthcare, the secure storage and transmission of heterogeneous medical images remain critical bottlenecks in telemedicine and cloud-based healthcare. Conventional multi-image encryption schemes still suffer from inherent limitations, including the independent processing of individual images, poor compatibility with varying image sizes, and reliance on fixed keys. To overcome these challenges, we propose a secure multi-image encryption scheme combining a memristive Hopfield neural network (MHNN) and the semi-tensor product (STP). Specifically, a six-dimensional MHNN exhibiting extreme multi-stability is constructed, capable of generating a large number of coexisting chaotic attractors with controllable positions, amplitudes, and topological structures. The pseudo-random sequences derived from these attractors pass rigorous randomness tests and provide an exceptionally large key space. Furthermore, a unified STP-based encryption paradigm is sysstematically developed, extending from single-image to multi-image scenarios. In multi-image applications, heterogeneous images are reconstructed into three dimensional tensors via zero-redundancy pixel stream stitching, enabling comprehensive encryption across images, spatial domains, and numerical ranges. Experimental results demonstrate near-ideal performance in aspects such as pixel correlation and information entropy, supporting large-scale medical image transmission while ensuring data integrity.

The memristor’s inherent memristive and nonlinear properties make it particularly well-suited for simulating synaptic connections in neural networks, inducing rich dynamical behaviors that are essential for understanding brain mechanisms and brain-like learning. In this paper, a new locally active non-volatile trigonometric memristor is constructed and coupled into the Hopfield neural network. The motion state of the memristive Hopfield neural network (MHNN) is influenced by the coupling strength, allowing it to exhibit periodic initial offset boosting behavior. Furthermore, the MHNN demonstrates unique multi-scroll attractor extension behaviors. The number of attractor scrolls increases continuously with parameters variations at fixed time intervals, although there is an upper limit. However, as simulation time extends, the number of attractor scrolls can grow indefinitely, with newly formed scrolls extending monotonically in both directions under different conditions. The MHNN is implemented using both analog circuits and the DSP platform. Eventually, a real-time image encryption scheme aimed at protecting medical image privacy is designed, supported by practical test. In particular, the scheme can be further applied to remote video medical protection, which can encrypt the treatment content.

Multimodal emotion recognition in conversation (MERC) has garnered substantial research attention recently. Existing MERC methods face several challenges: (1) they fail to fully harness direct inter-modal cues, possibly leading to less-than-thorough cross-modal modeling; (2) they concurrently extract information from the same and different modalities at each network layer, potentially triggering conflicts from the fusion of multi-source data; (3) they lack the agility required to detect dynamic sentimental changes, perhaps resulting in inaccurate classification of utterances with abrupt sentiment shifts. To address these issues, a novel approach named GraphSmile is proposed for tracking intricate emotional cues in multimodal dialogues. GraphSmile comprises two key components, i.e., GSF and SDP modules. GSF ingeniously leverages graph structures to alternately assimilate inter-modal and intra-modal emotional dependencies layer by layer, adequately capturing cross-modal cues while effectively circumventing fusion conflicts. SDP is an auxiliary task to explicitly delineate the sentiment dynamics between utterances, promoting the model's ability to distinguish sentimental discrepancies. GraphSmile is effortlessly applied to multimodal sentiment analysis in conversation (MSAC), thus enabling simultaneous execution of MERC and MSAC tasks. Empirical results on multiple benchmarks demonstrate that GraphSmile can handle complex emotional and sentimental patterns, significantly outperforming baseline models.