A complex-valued Hopfield network (CHN) is a multistate model of Hopfield neural network and has been applied to storage of image data. A two-level CHN (TLCHN) was proposed as a noise robust model against Gaussian noise. The disadvantage is that it has four times as many weight parameters as a CHN. To reduce the number of weight parameters, we propose a two-level quaternion-valued Hopfield network (TLQHN). A TLQHN has half weight parameters of TLCHN. A TLCHN improves the noise tolerance against Gaussian noise by the strength parameter, which is incorporated in the weight parameters. In a TLQHN, the strength parameter is hardly incorporated in the weight parameters, unlike a TLCHN. Instead, the strength parameter is incorporated in the weighted sum input. Computer simulations using image data show that the TLQHNs provide better noise tolerance than the models with same number of weight parameters.

Recent advancements in CNNs have significantly improved image classification, but the mechanisms behind feature extraction remain an open question. In particular, the relationship between human-defined semantic hierarchies and the visual feature hierarchies learned by CNNs remains unclear, and commonly used superclass labels (e.g., in CIFAR-100) may not align with the internal representation structure of these models. This study examines hierarchical grouping in CNNs using CIFAR-10 and CIFAR-100, focusing on superclass and fine-grained classification by utilizing t-SNE and CKA. The CKA results show that the first half of the layers of CNNs trained with the same dataset but different labels are similar. The results for t-SNE reveal that, in CIFAR-10, models group data by superclass before refining fine-grained categories in deeper layers. In CIFAR-100, the patterns are less consistent, with some superclasses following this trend while others do not, highlighting the limitations of manually defined superclass labels. To address this issue, a method is proposed to construct data-driven, feature-based superclasses using cosine similarity of latent representations and agglomerative hierarchical clustering. This approach leads to more coherent and interpretable clustering in t-SNE visualizations, particularly for VGG16. Furthermore, the proposed framework enables a comparative analysis of architectural differences, revealing that models with sequential feature extraction exhibit clearer hierarchical organization than architectures with residual connections.

Technology scaling and supply-voltage reduction increase the susceptibility of sequential circuits to radiation-induced soft errors. In particular, double-node upsets (DNUs) in the hold state and single-event transients (SETs) around clock transitions have emerged as major reliability concerns in nanoscale technologies. This work investigates two C-element-based latch (CEBLT) variants, CEBLT-1 and CEBLT-2, to improve robustness against both disturbance types. HSPICE simulations in a 15 nm FinFET process show that CEBLT-1 tolerates 60.0% of all possible DNU combinations, while the remaining failures are primarily caused by reverse propagation at the output stage. To mitigate this limitation, an enhanced latch structure, CEBLT-2, is introduced. The results demonstrate that the DNU tolerance coverage is increased to 88.9%, and further to 97.8% when CEBLT-2 is employed as the master stage of a master–slave flip-flop. Both latch designs also exhibit strong tolerance to input-node SETs around clock transitions. Compared with existing DNU-tolerant latch designs, the proposed latch family addresses hold-state DNU tolerance and clock-edge SET resilience within a unified architecture, while maintaining competitive area, delay, and power characteristics. These results indicate that C-element-based latch structures can be systematically enhanced to achieve improved soft-error robustness in advanced sequential circuits.

Random sample consensus (RANSAC), which is based on a repetitive sampling from a given dataset, is one of the most popular robust estimation methods. In this study, an energy-based model (EBM) for robust estimation that has a similar scheme to RANSAC, energy-based RANSAC (EB-RANSAC), is proposed. EB-RANSAC is applicable to a wide range of estimation problems similar to RANSAC. However, unlike RANSAC, EB-RANSAC does not require a troublesome sampling procedure and has only one hyperparameter. The effectiveness of EB-RANSAC is numerically demonstrated in two applications: a linear regression and maximum likelihood estimation.

In the self-dual nonlinear network lattice, the intrinsic localized mode (ILM) travels almost freely because of its very weak interaction with linear phonon modes, even when the nonlinear lattice is not integrable. In experiments, a propagating wave driver is used in the ring-shaped lattice to maintain the traveling ILMs by compensating for the unavoidable energy loss due to damping. The amplitude and velocity increase with the driver frequency F because of the positive nonlinearity of the lattice. Small steps are observed in the velocity curve when it is plotted as a function of the driver frequency. Between adjacent steps, the ratio (velocity)/(driver frequency) remains nearly constant, which implies that synchronization between the rotational motion of the ILM in the ring and the ILM vibration occurs. Simulations reveal that the appearance of the steps requires a faster velocity scaling as F^a (a ＞ 1, typically ∼ 1.4). The large damping at zero voltage and zero current – so called saturable damping – produces this velocity-boosting effect. The Fourier amplitude on the dispersion line (DL) of the traveling ILM is analyzed, and the shift of the DL spectrum explains the enhanced velocity.

Transfer learning with pre-trained models is essential in image classification, yet understanding why specific parameters enable effective adaptation remains limited. We investigate the role of Batch Normalization (BN) γ parameters in transfer learning, where training only BN parameters and the classifier achieves performance comparable to full fine-tuning. Through systematic experiments across multiple domains combining ablation studies, distribution analysis, overlap rate analysis, and correlation studies, we demonstrate that γ parameters function as adaptive feature selectors. Our findings reveal dataset-specific yet highly reproducible feature selection patterns, with quantitative evidence consistently supporting γ’s capability for deterministic task-specific feature selection from pre-trained models.

As one of the applications of X-ray computed tomography (X-ray CT) to geomaterials, rock CT images have been widely applied in earth and environmental sciences. However, the rock CT images have a low-resolution problem in the depth direction due to multiple causes such as physical characteristics of the rock core samples, geometric constraints of the imaging environments, and limitations in measurement in X-ray CT scanners. In this study, we propose a data-driven super-resolution based on generative modeling to improve the depth resolution of the rock CT images. Our proposed method solves the low-resolution problem as conditional generation by latent diffusion models which are a class of generative models. Latent diffusion models enable high-fidelity data generation by learning the time-reversed stochastic dynamics of a non-equilibrium diffusion process in a learned latent space, where latent representations are progressively transformed into Gaussian noise. In the proposed method, when we assume three consecutive images at different depth levels, a second image (an unobservable rock CT image) is generated from a first image and a third image (observable rock CT images) based on the diffusion mechanism. We verify the effectiveness of the proposed method by using actual rock CT images obtained in Oman Drilling Project, which is one of the international scientific research projects. The experimental results suggest that our proposed method can estimate the unobservable rock CT images more accurately than existing interpolation methods in the qualitative evaluation. In particular, the quantitative assessment using mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) as evaluation metrics for images shows significant improvement of the values calculated from the metrics.

Kernel-based learning methods such as Kernel Logistic Regression (KLR) can substantially increase the storage capacity of Hopfield networks, but the principles governing their performance and stability remain largely uncharacterized. This paper presents a comprehensive quantitative analysis of the attractor landscape in KLR-trained networks to establish a solid foundation for their design and application. Through extensive, statistically validated simulations, we address critical questions of generality, scalability, and robustness. Our comparative analysis shows that KLR and Kernel Ridge Regression (KRR) exhibit similarly high storage capacities and clean attractor landscapes under typical operating conditions, suggesting that this behavior is a general property of kernel regression methods, although KRR is computationally much faster. We identify a non-trivial, scale-dependent law for the kernel width γ, demonstrating that optimal capacity requires γ to be scaled such that γN increases with network size N. This finding implies that larger networks require more localized kernels, in which each pattern’s influence is more spatially confined, to mitigate inter-pattern interference. Under this optimized scaling, we provide clear evidence that storage capacity scales linearly with network size (P ∝ N). Furthermore, our sensitivity analysis shows that performance is remarkably robust with respect to the choice of the regularization parameter λ. Collectively, these findings provide a concise set of empirical principles for designing high-capacity and robust associative memories and clarify the mechanisms that enable kernel methods to overcome the classical limitations of Hopfield-type models.

Kernel-based learning methods can dramatically increase the storage capacity of Hopfield networks, yet the dynamical mechanisms behind this enhancement remain poorly understood. We address this gap by combining a geometric characterization of the attractor landscape with the spectral theory of kernel machines. Using a novel metric, Pinnacle Sharpness, we empirically uncover a rich phase diagram of attractor stability, identifying a Ridge of Optimization where the network achieves maximal robustness under high-load conditions. Phenomenologically, this ridge is characterized by a Force Antagonism, in which a strong driving force is counterbalanced by a collective feedback force. We theoretically interpret this behavior as a consequence of a specific reorganization of the weight spectrum, which we term Spectral Concentration. Unlike a simple rank-1 collapse, our analysis shows that the network on the ridge self-organizes into a critical regime: the leading eigenvalue is amplified to enhance global stability (Direct Force), while the trailing eigenvalues remain finite to sustain high memory capacity (Indirect Force). Together, these results suggest a spectral mechanism by which learning reconciles stability and capacity in high-dimensional associative memory models.

In this paper, we aim to show the usefulness of contextual information in distributed representations for the genre classification of modern Japanese literary works. Our previous studies have demonstrated that high-accuracy genre classification can be achieved by using distributed semantic representations generated by the Continuous Bag-of-Words (CBOW) model. Meanwhile, because BERT can acquire distributed semantic representations that vary according to context, it is expected to enable more precise semantic representations and, consequently, more accurate genre classification. Therefore, the purpose of this study is to compare the genre classification performance of BERT and the CBOW model based on their distributed semantic representations, and to clarify the differences in the representations obtained by the two models. The experimental results show that in the “novel vs. poetry” task, where the text forms differ greatly, both models achieved a high accuracy of 99%. On the other hand, in the “novel vs. essay” task, where the text forms are similar, the 500-dimensional BERT model achieved the highest accuracy of 92.25%, outperforming the CBOW model. These results indicate that the contextual representations acquired by BERT capture subtle differences in word usage, thereby contributing to improved classification performance in genre pairs with similar sentence structures.

Lightweight CNNs such as MobileNetV2, ShuffleNetV2, and ResNet18 exhibit irregular computation patterns—including depthwise separable chains, residual shortcuts, and branch-merge structures that form complex dependency graphs. On PE-array accelerators with limited on-chip buffers, conventional layer-wise and heuristic pipelined scheduling fails to exploit these dependencies, resulting in synchronization stalls, weight-bank contention, and redundant computation from spatial tile overlap. We proposed a structure-aware tile-level scheduling framework that explicitly models fine-grained dependencies under fixed hardware resources. The framework constructs an operator dependency graph into a tile-level DAG, and enables cross-layer tile pipelining through dependency-preserving scheduling. A PPO-based scheduler is trained to minimize makespan while penalizing weight conflicts and spatial overlap redundancy. Experimental results show that the proposed method consistently outperforms layer-sequential and layer-pipelined baselines, achieving up to 83% latency reduction over layer-wise execution and up to 10% improvement over pipelined scheduling. Average PE utilization is improved by 6–10% without increasing hardware parallelism, indicating that performance gains arise from dependency-aware tile ordering and reuse-aware PE allocation rather than architectural scaling.

Extracting nonlinear neuronal dynamics is one of the important subjects in neuroscience. However, we can access only limited and partially observable low-dimensional data with noise in many situations. In this study, we propose a data-driven method for simultaneously estimating membrane potential dynamics and calcium dynamics from partially observable noisy time-series data. We derive a sequential Monte Carlo method for estimating multi-dimensional neuronal dynamics from a conductance-based spiking neuron model. Furthermore, we derive an expectation-maximization algorithm for estimating membrane conductances by reflecting both membrane potential dynamics and calcium dynamics. Using the proposed method, we show that the proposed framework is effective for extracting neuronal membrane potential and calcium dynamics simultaneously.

With the development of information technology in recent years, the analysis of various data types has become increasingly important. In particular, analysis methods for relational data, in which the strength of relationships between objects is represented numerically, have attracted considerable attention. Conventional relational data analysis methods focus solely on the relationships between objects. However, in recent years, it has become common to obtain attribute information for each object in addition to relational data, necessitating the development of methods that integrate and analyze these data jointly. Although latent structure analysis methods for relational data with attribute information have been proposed, they exclusively target single-domain data, with no extensions to multiple domains having been conducted. Based on this background, this study proposes a method to extend existing relational data analysis methods with attribute information from single-domain to multiple domain settings. Furthermore, unlike conventional methods that rely on iterative algorithms, we propose a non-iterative estimation method for latent variables and parameters by combining multivariate analysis techniques, as well as a method to enhance the applicability to diverse data types. The effectiveness of the proposed method is demonstrated through experiments using both synthetic and real data.

Distributed edge learning degrades under heterogeneous, evolving tasks and intermittent client participation, where naive peer-side averaging causes model drift and costly communication. We propose task-similarity-aware, confidence-gated partial model fusion that integrates updates only from task-aligned, reliable peers. We introduce JSDCosNet for collaborator selection by combining Jensen-Shannon divergence and cosine similarity, refined via asymmetric feedback collaboration. We employ adaptive multi-donor fusion to address non-stationary events. Our framework partially integrates common-task components weighted by the expertise score, restricting uncertain updates. Evaluations on CIFAR-10 improve target-class test accuracy and training stability in decentralized settings, while achieving a 4.5× reduction in per-update communication cost.

This paper proposes the Inertial Cubic regularized Newton (ICN) method, which incorporates inertial terms into the Cubic regularized Newton (CN) framework to accelerate convergence while preserving numerical stability. With the rapid growth of data-driven applications, conventional Newton-type methods have increasingly suffered from instability caused by indefinite or ill-conditioned Hessian matrices. The CN method addresses these issues by employing cubic regularization, ensuring global convergence to stationary points under standard assumptions and improved robustness. However, its convergence speed can still be limited in practice. To overcome this limitation, the proposed ICN method introduces inertial terms, originally developed for accelerating first-order optimization methods, into the second-order CN framework. This integration enhances convergence efficiency without compromising solution accuracy or stability. The effectiveness of the proposed method is demonstrated through extensive computational experiments on benchmark function optimization, logistic regression, and pattern classification problems. The results show that the ICN method consistently achieves faster convergence and improved robustness compared to existing second-order optimization algorithms.

The rapid evolution of Large Language Models (LLMs) has necessitated robust technologies to distinguish AI-generated text from human-written content. However, existing detection methods often rely on semantic content words, causing significant performance degradation when applied to Out-of-Distribution (OOD) domains where the topic or writing style differs from the training data. To address this limitation, this paper proposes a domain-agnostic detection framework that focuses on the structural degeneration observed in Function Words (FW). We hypothesize that while content words vary by topic, the probabilistic maximization bias of AI models manifests universally as excessive typicality in function word usage. We introduce two approaches using the RoBERTa encoder: FW-RoBERTa, which utilizes averaged hidden state vectors of function words, and FW-PLL, which utilizes function-word-restricted Pseudo-Log-Likelihood scores. By filtering out content words, our method isolates the syntactic structure from explicit semantic content. Extensive experiments across multiple domains (arXiv, XSum, WritingPrompt) and generative models, ranging from GPT-3.5 Turbo to the state-of-the-art GPT-5.1 and Gemini 3 Pro, demonstrate that our approach maintains high accuracy in OOD environments where conventional baselines fail. These results confirm that context-aware function word features capture the intrinsic, invariant differences between human and AI text.

Visual Place Recognition (VPR) under severe environmental changes remains a fundamental challenge for autonomous roboticsin real-world environments. This task can be interpreted as associative memory retrieval from noisy queries, but classical models suffer from limited capacity and sensitivity to pixel-level variations. We address this by integrating Modern Hopfield Networks with DINOv3, a self-supervised Vision Transformer that provides robust semantic representations. The primary aim of this study is not to maximize VPR accuracy itself, but to investigate whether an energy-based associative memory can be realized on the latent space of a foundation model, using VPR as a challenging real-world testbed. Place recognition is formulated as energy minimization in a semantic latent space, where stored scenes act as attractors. Experiments on the Transient Attributes Database across four seasons show that the proposed method significantly outperforms pixel-based baselines, even under extreme domain shifts. We further analyze the retrieval dynamics and the effect of the inverse temperature parameter β on attractor stability.

The Optimizer based on Spiking Neural-oscillator Networks (OSNN) is a deterministic swarm intelligence algorithm inspired by Particle Swarm Optimization (PSO). OSNN consists of multiple particles, each composed of multiple spiking oscillators. Each spiking oscillator independently searches a one-dimensional solution space and is coupled to its neighboring oscillator through a Ring 1-way network. In our previous studies, OSNN demonstrated improved performance compared with the classical stochastic PSO, and the effectiveness of the Ring 1-way network was clarified. In the conventional OSNN, the Ring 1-way network is incorporated into spiking oscillators that explore the same dimensional solution space, which is an intra-dimensional coupled network. As a result, spiking oscillators interact with those belonging to other particles, whereas spiking oscillators within the same particle do not interact with one another. Although the Ring 1-way network affects the search dynamics and firing rates of spiking oscillators, the effectiveness of incorporating this network into spiking oscillators within the same particle has not yet been clarified. In this study, we propose an inter-dimensional coupled network by incorporating the Ring 1-way network into spiking oscillators within each particle. Furthermore, we combine the inter-dimensional coupled network with the intra-dimensional coupled network in OSNN, forming the mesh network. The proposed networks increase the firing rates of spiking oscillators in the swarm, and the mesh network, in particular, achieves higher firing rates than the intra-dimensional and the inter-dimensional coupled networks. Numerical experiments are conducted to evaluate the effectiveness of the proposed networks in OSNN by comparing them with the Optimizer with Uncoupled Spiking Neurons (OUSN), OSNN with the intra-dimensional coupled network, and the classical PSO. The simulation results reveal that higher firing rates tend to be associated with better performance when incorporating the Ring 1-way network into OSNN, and OSNN with the mesh network achieves the best performance among the comparison methods.

In this study, an ergodic sequential logic (ESL) cochlear model is designed for the purpose of reproducing nonlinear sound processing functions of mammalian cochleae. It is demonstrated that the designed cochlear model is capable of reproducing typical nonlinear sound processing functions of the mammalian cochleae, including combination tone generation as well as low-side and high-side suppressions. Furthermore, the designed cochlear model is implemented on a field programmable gate array and its operations are validated by experiments. It is then shown that the designed cochlear model is much more hardware efficient compared to a conventional ordinary differential equation cochlear model.

In this paper, we investigate a dynamical personal authentication method based on natural facial expression changes. Fingerprint authentication and conventional face authentication are static pattern-based methods, and their vulnerability to imitation using artificial objects is a significant security concern. In contrast, it is difficult to imitate human dynamic characteristics, and incorporating dynamic information into face authentication can lead to a more robust authentication method. Therefore, the purpose of this paper is to realize a dynamical personal authentication method based on natural facial expression changes. We generated difference data of facial feature vectors between two frame images constituting a video, and by using vector components in which statistically significant differences were observed, personal classification was possible with an average accuracy of 97.73%. Furthermore, using the best-performing model as the authentication module, we conducted dynamical personal authentication experiments and achieved FAR = 0% for 7 out of 13 subjects.

Echo state networks (ESNs) can achieve high prediction accuracy for time series; however, their performance is highly sensitive to the choice of hyperparameters. Although the hyperparameter selection depends on the characteristics of target systems, especially the autocorrelation structure of time series, the relationship between the hyperparameter selection and the characteristics of target systems remains unclear. Therefore, in this paper, we adjust the time scale of the target time series to clarify the relationship between their autocorrelation structure and suitable hyperparameters of ESNs in prediction tasks. We modify the time scale while preserving essential dynamics by introducing the concept of decorrelation time. In numerical experiments, we performed predictions using an ESN for three chaotic time series with adjusted time scales to investigate the effects of the time scale on the hyperparameter selection in ESNs. The results show that by the adjustment to the same time scale of the time series, parameter regions with high prediction accuracy exhibit similar structures regardless of the target time series. In particular, when predicting time series of longer time scales, high prediction accuracy was obtained in the range in which the spectral radius exceeds unity. These results emphasize the importance of determining the hyperparameters of ESNs based on the time scale of the target time series and provide universal guidelines to select the hyperparameters of ESNs.

Recent advances in deep generative models have enabled the synthesis of highly photorealistic images, raising concerns about image authenticity and misinformation. While generated image detection has been widely studied for faces and objects, landscape images remain challenging due to their complex semantic structure and strong dependence on global factors such as illumination and perspective. In this study, we investigate the characteristics of AI-generated landscape images and analyze the decision-making basis of convolutional neural networks (CNNs) from an explainable AI perspective. A carefully controlled dataset of real and generated landscape images across ten categories is constructed, and a baseline CNN using only RGB inputs is evaluated. SHAP and Grad-CAM analyses reveal that the model relies primarily on local texture irregularities and smoothness artifacts. Robustness experiments using masking and retraining demonstrate that detection remains stable under blurring and edge removal but degrades significantly when local visual information is completely suppressed. Finally, a multimodal model integrating semantic segmentation masks with RGB inputs is proposed, achieving improved accuracy and interpretability by explicitly leveraging region-level structural information.

In this study, we investigated how differences in arm shaking actions and user attentiveness influence the degradation of personal authentication accuracy over time. Our previous research has shown that variations in arm shaking actions lead to differences in authentication accuracy after a one-month period. These findings suggest that the naturalness of the arm shaking action may be associated with the ability to sustain authentication performance over time. To further explore this relationship, we investigated the effects of both action style and user attentiveness on long-term authentication accuracy. The results indicate that arm shaking actions performed with the smartphone screen oriented perpendicular to the ground achieved low FRR and FAR performance, even during video viewing and after a one-month interval. This suggests that such shaking actions are not consciously controlled but instead are natural movements, and are therefore effective in maintaining high authentication performance over extended periods.

In recent years, neural network technology has been widely used in various situations. However, as neural networks have become larger and deeper, various issues have arisen, including increased circuit scale, power consumption, and latency. To address such challenges, model compression methods such as binary neural networks and implementation methods on FPGAs or ASICs have been proposed. In this study, we focus on FPGA implementation of binary neural networks and propose a more efficient implementation method than conventional methods by designing the activation functions based on majority logic.

Chaotic search has demonstrated promising performance for solving combinatorial optimization. However, its effectiveness is highly sensitive to parameter settings of chaotic neural networks, and empirical tuning often fails to maintain stable performance across different problems and searching stages. To address this issue, we propose a chaotic search method with particle swarm optimization, a learning-based adaptive tuning method that integrates particle swarm optimization into the chaotic search framework. Within this framework, particle swarm optimization serves as an online learning mechanism that dynamically optimizes key parameters, enabling the automatic regulation of neural excitation during the search process. Comparative experiments on capacitated vehicle routing problems reveal that the chaotic search method with particle swarm optimization has better solution quality and higher robustness compared with conventional chaotic search and feedback-based tuning methods. These results confirm the efficacy of swarm-based online parameter learning for enhancing search performance of chaotic search.

Anticipating soccer penalty kick direction is crucial for goalkeepers because reaction time alone is often insufficient after ball contact. This study examines whether kick direction (right vs. left) can be predicted from pre-impact kinematics estimated by markerless pose tracking. Four participants performed left- and right-targeted penalty kicks, and 2D coordinates of 25 joints were extracted using OpenPose. Echo State Networks (ESNs) were trained in a joint-wise manner using single-joint coordinate time series, with a ridge-regression readout. High prediction performance as achieved under kicker-specific conditions, whereas cross-subject generalization was limited. A decision-basis analysis using the decision-axis projected update (DAPU) further revealed time intervals that consistently contributed to decision formation.

Wireless Brain-Inspired Computing (WiBIC) has been proposed as a neuromorphic AIoT(AI + IoT) platform that integrates wireless communication and spiking neural computation in a distributed, serverless manner. While prior studies have demonstrated the feasibility of WiBIC through simple single-task examples, its fundamental capability as a practical AIoT platform has not yet been fully established. This paper evaluates the foundational capabilities and practical viability of WiBIC through the development of two representative systems. First, an operant conditioning learning system is implemented to demonstrate reward-driven behavioral adaptation via distributed reinforcement learning without centralized control. Second, an in-room occupancy estimation system is developed, in which multiple ambient sensors-illuminance, human-detection, and current-are treated as multimodal inputs. Using reservoir computing with delay adaptation, the system performs privacy-preserving, serverless edge inference to estimate occupancy levels. Through these systems, we demonstrate that WiBIC can consistently unify computation and communication under spike-based processing, enabling distributed learning and inference on AIoT devices. The results provide concrete evidence that WiBIC possesses the fundamental capabilities required of a practical neuromorphic AIoT platform.

Elucidating the phenomenon of walking synchronization on suspension bridges requires investigation from both physical experimental and mathematical modeling perspectives. In this paper, we propose an experimental system and its corresponding mathematical model, consisting of a laterally oscillating robot with leg extension/retraction capabilities and a suspension bridge. The model successfully reproduces the characteristic behaviors observed in the experimental system. These findings suggest that the proposed model can be a useful framework for elucidating the mechanism of walking synchronization.

Graph neural networks (GNNs) are some of the architectures used in physics-informed machine learning to model graph-structured systems. Such physical systems are not solely governed by their topological structure; boundary conditions are important to determine system dynamics. However, boundary conditions are difficult to incorporate into conventional GNNs because a graph and its Laplacian assume homogeneous Neumann boundaries implicitly. We propose a graph representation and its Laplacian with boundary conditions inspired by the concept of ports in circuit theory. Furthermore, we develop a GNN framework that accounts for boundary effects. We demonstrate the effectiveness of our GNN through electrical circuit datasets.

Constructing adjacency relationships between surface patches is a fundamental step in surface reconstruction for reverse engineering, using point cloud data. Conventional approaches rely on polygonal meshes, often requiring hole filling and manual preprocessing of scanned data. We propose a simple method for constructing adjacency graphs directly from segmented point clouds, without explicit mesh generation. Point cloud segments are embedded into a spherical latent space by an autoencoder, and adjacency is estimated by partitioning the sphere surface. The method was evaluated on synthetic datasets and measured point clouds acquired using 3D scanning. The results show that accurate adjacency graphs can be obtained from point clouds alone. Ablation studies validated the spherical constraint, latent dimensionality, and network topology used in this framework.

Indoor localization using smartphones remains challenging due to sensor drift and unstable wireless signals. This paper proposes a lightweight fusion framework combining an improved Pedestrian Dead Reckoning (PDR) algorithm with local Bluetooth Low Energy (BLE) fingerprinting. A machine-learning-based step length model and an online yaw correction strategy are introduced to enhance PDR robustness. Unlike conventional fingerprinting methods requiring dense deployment, a sparse anchor zone strategy is proposed, where BLE beacons are placed only at key locations to periodically correct PDR drift. Experiments on a public dataset and real-world tests demonstrate effective drift reduction with low deployment and computational cost.

Recent deep neural classifiers may operate under a dimension-deficient setting, where the feature dimension is insufficient relative to the number of classes and the symmetric geometry predicted by Neural Collapse cannot be realized. In this study, we focus on one resulting geometric structure, the Island Class Structure, which forms a bounded internal decision region near the origin. Using theoretical expressions of decision boundaries, we derive its emergence conditions and validate them experimentally on MNIST and CIFAR-10. The results show that this structure reproducibly appears as an intermediate geometric solution under severe low-dimensional constraints.

The spatiotemporal learning rule (STLR) updates the synaptic weights based on the coincidence among inputs to a neuron, enabling it to discriminate subtle spatiotemporal differences, in contrast to the Hebbian learning rule. In this study, we evaluated the learning performance of STLR by varying the learning coefficient and the time constant of the time history in coincidence. To quantitatively evaluate learning performance, we improved the multistage template-matching method. The results demonstrated that introducing an appropriate time constant enhances context retention and mitigates forgetting during spatiotemporal pattern learning. Furthermore, we used the Modified National Institute of Standards and Technology dataset as actual spatiotemporal patterns. In conclusion, the time history of coincidence is essential for effectively capturing and learning the contextual structures in spatiotemporal patterns.

Multimodal inter-spike interval (ISI) distributions are key characteristics of neuronal information processing. This phenomenon suggests that subthreshold oscillations in neurons exhibit isochronous characteristics, maintaining a constant period independent of amplitude. However, conventional piecewise-constant (PWC) models lack isochrony, and fail to reproduce multimodal ISI distributions. To address this, we propose a novel method that introduces isochrony into PWC dynamics by state-dependent modulation of the vector field. Our model successfully reproduces multimodal ISI distributions while retaining analytical tractability. We derive the theoretical ISI distribution using a one-dimensional return map and validate the results through circuit experiments.

Asynchronous pulse code multiple access (APCMA) is a pulse-based scheme designed for massive IoT that offers robustness against signal collisions compared with other LPWA systems. However, APCMA faces critical challenges such as performance degradation due to pulse loss and false detection, and difficulty in tuning pulse detection parameters under low-SNR. This paper proposes and implements a lightweight machine-learning-based pulse detection method and evaluates it in high-attenuation and high-density environments. Experimental results demonstrate more than 10 % improvement in decode success rate at SNR of −31 dB without manual threshold tuning and confirm the effectiveness of the proposed method in wireless environments.

Adaptive wireless resource selection is crucial in dynamic radio environments with rapidly changing traffic. Previous studies demonstrated ultra-fast reinforcement learning using laser chaos dynamics and the tug-of-war (TOW) principle, achieving GHz-speed solutions for multi-armed bandit (MAB) problems with time-varying rewards. Another photonic approach employs ring-configured laser networks that scale well and remain largely insensitive to the number and arrangement of options, but it struggles in dynamic settings because its value estimates retain outdated reward information. This paper introduces a decay factor to discount past rewards, improving responsiveness and enabling ultra-fast adaptation. Simulations for Wi-Fi channel selection confirm convergence to optimal channel assignment.

Ultra-massive multiple-input multiple-output (UM-MIMO) systems are expected to support extremely dense user connectivity in beyond fifth-generation (5G) and sixth-generation (6G) wireless communications. In such systems, conventional plane-wave-based beamforming approaches become inadequate due to near-field propagation characteristics and the resulting spatial non-stationarity of electromagnetic fields. Therefore, a directivity optimization framework that explicitly accounts for near-field propagation effects and spatial degrees of freedom is required. This paper proposes a beamforming framework for multi-user MIMO (MU-MIMO) systems based on spherical mode expansion (SME). By formulating the beam design problem in the spherical modal domain, user-specific optimal spherical mode coefficients (SMCs) are analytically derived based on the eigenmode of the spherical mode correlation matrix. The proposed framework combines SME-based directivity optimization with DFT-based beamforming, enabling efficient spatial multiplexing while reducing inter-user correlation. The influence of array parameters, including array radius, element density, and user spatial distribution, is systematically evaluated through numerical simulations using a hemispherical antenna array. Simulation results demonstrate that the proposed SME-based method consistently achieves higher channel capacity and stronger robustness against increasing user density compared with conventional zero-forcing-based schemes. The results further clarify that satisfying the spatial resolution requirements in the spherical modal domain is essential for fully exploiting higher-order spatial modes. These findings indicate that SME provides a physically grounded and effective foundation for MU-MIMO beamforming and directivity optimization in future UM-MIMO and 6G wireless communication systems.

Spiking neural networks are well-suited for the efficient utilization of neuromorphic devices. Spatial navigation is one possible implementation of those devices. In the brain, place cells and grid cells contribute to spatial representation. The formation of grid cells by spiking neurons enables spatial representation in neuromorphic devices. In this study, we propose a method to convert an artificial neural network for the grid cell into a spiking neural network. Our spiking neural network consists of integrate-and-fire neurons with inputs from the moving speed and head direction, and the neurons’ outputs form firing-rate maps of place cells. Experimental results showed that grid patterns were observed in the spiking neural network, and the self-position from one’s movement was successfully determined.

It has been well-known that many real networks are scale-free (SF) but extremely vulnerable against attacks. We investigate the robustness of connectivity and the lengths of the shortest loops in randomized SF networks with realistic exponents 2.0 < γ ≤ 4.0. We show that smaller variance of degree distributions leads to stronger robustness and longer average length of the shortest loops, which means the existing of large holes. These results will provide important insights toward enhancing the robustness by changing degree distributions.

Koopman theory has become widely used for predictions and analysis of nonlinear systems. Especially, kernel extended dynamic mode decomposition (kEDMD) embeds nonlinear dynamics in a reproducing-kernel Hilbert space. However, its Gram matrix grows with every snapshot, blocking online use when parameters drift. We present a low-rank approximation-based kEDMD algorithm that swaps out arbitrary old snapshots, inserts new ones, and updates the eigendecomposition with rank-two corrections, keeping both memory and cost fixed. The method efficiently tracks evolving systems on the van der Pol oscillator and adapts to unseen data and slow coefficient changes. In addition, it efficiently tracks a 50-point discretized Korteweg-de Vries flow and preserves predictive accuracy.

Convective features, represented here as warm bubble-like patterns, reveal essential high-level information about how short-term weather dynamics evolve within a high-dimensional state space. In this paper, we introduce a data-driven framework that uncovers transient dynamics captured by Koopman modes responsible for these structures and traces their emergence, growth, and decay. Our approach applies the sparsity-promoting dynamic mode decomposition to weather simulations, yielding a few number of selected modes whose sparse amplitudes highlight dominant transient structures. By tuning the sparsity weight, we balance reconstruction accuracy and model complexity. We illustrate the methodology on weather simulations, using the magnitude of velocity and vorticity fields as distinct observable datasets. The resulting sparse dominant Koopman modes capture the transient evolution of bubble-like pattern and can reduce the dimensionality of weather system model, offering an efficient surrogate for diagnostic and forecasting tasks.

The concept of neural codes is instrumental in understanding how the nervous system represents and processes information. However, traditional approaches to neural coding rely almost exclusively on observer-dependent descriptions of neural activity, which may lead to potentially ambiguous or inconsistent interpretations. In this paper, we propose a geometric approach to quantifying the information content of spike sequences from a neural-centric perspective that is distinctly independent of any observer-based assumptions. We benchmark our method by conducting spike-train classification experiments on both synthesized and real datasets and comparing it with widely used spike-train metrics. The comparison results validate the discriminative power of our model and its effectiveness in quantifying and visualizing the coded information in spiking activity across a variety of neural processes. In particular, our approach may unify both rate and temporal coding schemes, thereby providing a new geometric interpretation that contributes to resolving the renowned debate regarding the nature of neural coding.

In this paper, we propose a randomized assignment technique of real and imaginary constellations for orthogonal frequency division multiplexing (OFDM) signals using binary phase shift keying (BPSK) over memoryless nonlinear channels to improve bit error rate (BER) performance. The BER performance of the proposed method is theoretically analyzed and compared with that of the conventional BPSK-OFDM scheme, in which all active subcarriers use the same constellation on the in-phase axis. The effectiveness of the proposed method and the validity of the theoretical analysis are verified through computer simulations and experiments using actual hardware setups.