This paper investigates the application of next-generation reservoir computing (NGRC) to channel equalization in coherent optical communication systems. Reservoir computing (RC) has attracted attention as a low-cost machine learning framework for time-series processing, and its complex-valued extension (CVRC) has shown improved performance for optical channel equalization. However, NGRC, which eliminates recurrent connections and enables faster computation, has not yet been explored for optical channel equalization. In this study, we propose an NGRC-based equalizer and its complex-valued counterpart (CVNGRC) for coherent optical communication systems. We design input features and nonlinear polynomial expansions tailored for channel equalization and evaluate their effectiveness through numerical simulations. The performance of the proposed methods is compared with conventional RC and CVRC in terms of symbol error rate (SER), training time, and inference time. Simulation results demonstrate that real-valued NGRC (RVNGRC) significantly reduces computational cost while maintaining equalization performance comparable to that of real-valued RC (RVRC). Furthermore, CVRC achieves the best SER performance due to its ability to directly process complex-valued signals, whereas RVNGRC achieves a balance between computational efficiency and accuracy. These findings clarify the performance–complexity trade-offs among RC-based equalizers and highlight the potential of RVNGRC for high-speed optical communication systems.

Conventional image denoising models trained with mean squared error achieve high signal fidelity but often produce over-smoothed outputs with weak high-frequency textures. To address this, we propose a perceptual quality-aware fine-tuning framework for NAFNet, in which an MSE-pretrained model is fine-tuned using a hybrid objective combining MSE, LPIPS, and DISTS. We further introduce post-training model merging via linear interpolation between the MSE baseline and the perceptually fine-tuned model, enabling continuous control of the signal-perception trade-off without retraining. Experiments on SIDD and MIDD across two sensor domains show that perceptual fine-tuning substantially improves perceptual metrics with limited PSNR degradation, and that intermediate merge ratios can simultaneously improve perceptual quality while maintaining or slightly improving PSNR relative to the baseline.

Image inpainting methods based on latent diffusion models offer broad generalization and text-conditioned generation, but suffer from two limitations in zero-shot settings. First, boundary discontinuities arise from approximate latent-space projection. Second, obtaining suitable text prompts from masked inputs is inherently difficult. We propose a zero-shot framework that addresses both issues by extending a latent diffusion-based inpainting method with two modules. A large vision-language model generates degradation-robust prompts under instructions to ignore masked regions, and a pixel-space diffusion model is applied as post-processing to boundary regions to suppress discontinuities. Experiments on BSDS500 and DIV2K show that the proposed method outperforms all zero-shot baselines in FID and achieves competitive results against task-specifically trained methods without any task-specific training.

Reservoir computing has shown strong performance in time-series prediction, but systems with both parameter variation and periodic forcing remain challenging to model. We propose a reservoir computing framework that incorporates parameter-related and periodic external inputs, enabling a unified treatment of nonstationarity and nonautonomy. Experiments with the Duffing system show consistently improved prediction performance over traditional reservoir computing for trained, untrained, and nearby extrapolative parameter regimes. In addition, a snow-depth dataset is used as a real-world example of such dynamics to examine whether the same framework can capture year-to-year variability and seasonal structure.

The two-dimensional cellular automaton model proposed by D. A. Young can form the Turing patterns. Previous studies have reported that rare patterns from the original model appear more frequently in the extended Young’s model. However, this increase in frequency has not been quantitatively evaluated. In this paper, we formulate the emergence probability of these patterns based on the size of their basins of attraction. To estimate this size, we use an automatic classifier for two-dimensional Turing patterns and a stratified sampling method. The results quantitatively confirm that the extended model yields a higher emergence probability for these patterns.

Gene network simulators have the potential to accelerate large-scale parameter exploration toward drug discovery. In this paper, a systematic design method for a hardware-efficient ergodic sequential logic (ESL) gene network model is proposed. The ESL gene network model introduces a rotatable lattice in the state space, and the designed ESL model reproduces representative dynamical behaviors of a target gene network model. The designed ESL model is also implemented on an FPGA and shown to require fewer circuit elements and lower energy consumption than a customized DSP-based implementation. These results provide a practical basis for future hardware-based gene network simulators.

This paper investigates the nonlinear dynamics and phase transitions in power packet network connected with routers, conceptualized as macroscopic information-ratchets. In the emerging paradigm of cyber-physical energy systems, the interplay between stochastic energy fluctuations and the thermodynamic cost of control information defines fundamental operational limits. We first formulate the dynamics of a single router using a Langevin framework, incorporating an exponential cost function for information acquisition. Our analysis reveals a discontinuous (first-order) phase transition, where the system adopts a strategic abandon of regulation as noise intensity exceeds a critical threshold D_c. This transition represents a fundamental information-barrier inherent to autonomous energy management. Here, we extend this model to network configurations, where multiple routers are linked through diffusive coupling, sharing energy between them. We demonstrate that the network topology and coupling strength significantly extend the bifurcation points, with collective resilient behaviors against local fluctuations. These results provide a rigorous mathematical basis for the design of future complex communication-energy network, suggesting that the stability of proposed systems is governed by the synergistic balance between physical energy flow and the thermodynamics of information exchange. It will serve to design future complex communication-energy networks, including internal energy management for autonomous robots.

Recent developments in hardware, such as photonic integrated circuits and optical devices, are driving demand for research on constructing machine learning architectures tailored for linear operations. Hence, it is valuable to explore methods for constructing learning machines with only linear operations after simple nonlinear preprocessing. In this study, we propose a framework to extract a linearized model from a pre-trained neural network for classification tasks by integrating Koopman operator theory with knowledge distillation. Numerical demonstrations on the MNIST and the Fashion-MNIST datasets reveal that the proposed model consistently outperforms the conventional least-squares-based Koopman approximation in both classification accuracy and numerical stability.

Identifying governing equations of nonlinear dynamical systems from data is challenging. While sparse identification of nonlinear dynamics (SINDy) and its extensions are widely used for system identification, operator-logarithm approaches use the logarithm to avoid time differentiation, enabling larger sampling intervals. However, they still suffer from the curse of dimensionality. Then, we propose a data-driven method to compute the Koopman generator in a low-rank tensor train (TT) format by taking logarithms of Koopman eigenvalues while preserving the TT format. Experiments on 4-dimensional Lotka–Volterra and 10-dimensional Lorenz–96 systems show accurate recovery of vector field coefficients and scalability to higher-dimensional systems.

Although AI technologies have advanced rapidly and become widely adopted, the computational demands associated with increasing model size have become a major challenge, making energy efficiency an urgent concern for large-scale models. In response, this paper focuses on classifiers based on Hyperdimensional Computing (HDC) because of their high computational efficiency and lightweight model structure. However, the internal mechanisms of HDC-based classifiers remain insufficiently characterized, and their inference accuracy is often too low for practical deployment. Therefore, this study structurally analyzes HDC-based classifiers and evaluates two directions for characterizing and improving their representations: (1) channel-wise encoding with alternative position-vector assignments, and (2) the identification of informative dimensions followed by retraining and bit-expanded inference. These investigations clarify the behavior of HDC classifiers and demonstrate improved inference accuracy while preserving lightweight operations.

Deep super-resolution (SR) models frequently suffer from performance drops when real-world degradations differ from training assumptions. This degradation mismatch poses a significant challenge for practical applications. To diagnose the internal mechanisms of this failure, we propose a framework that visualizes the frequency response of SR models using Modulation Transfer Function (MTF) analysis. By inputting direction-specific sinusoidal patterns and analyzing the output spectra, we construct Angular MTF Maps that characterize model behavior. Our experiments reveal three key insights: (1) models trained on fixed kernels (e.g., Bicubic) implicitly learn rigid inverse filters, exhibiting distinct periodic artifacts; (2) degradation mismatch disrupts these learned priors, resulting in interpretable failure modes such as isotropic cutoff or directional attenuation; and (3) real-world SR models trained with diverse degradations (e.g., Real-ESRGAN) maintain consistent and stable frequency responses. We conclude that bridging deep learning with optical frequency analysis is essential for designing robust SR architectures that align with physical imaging realities.

This paper addresses the challenge of deploying Transformer models on edge devices, where the quadratic computational and memory costs of the attention mechanism limit long-sequence processing. We propose an efficient hardware architecture based on low-rank approximation via the Nyström method. To optimize performance, we introduce hierarchical optimizations including operation integration, memory access control, and pipelining. Theoretical evaluations confirm that our architecture suppresses the growth in internal memory usage and stabilizes memory bandwidth as sequence length increases, enabling highly scalable and efficient processing under severe resource constraints.

There is a growing expectation for enabling beyond-visual-line-of-sight (BVLOS) flight of unmanned aerial vehicles (UAVs) to support agricultural activities in mountainous regions with steep terrain. To achieve that, it is essential to estimate the operational range of UAVs within which stable wireless communication with the control terminals can be maintained. Existing studies propose methods for estimating wireless communication quality around control terminals by utilizing detailed radio propagation simulations based on ray-tracing techniques, but the computational cost increases exponentially when simulating the impact of terrain with complex non-geometric structures (e.g., vegetation in mountainous areas). To address this issue, this study proposes a novel method leveraging machine learning techniques to derive accurate estimates of wireless communication quality based on detailed three-dimensional terrain structures observed via remote sensing capabilities mounted on UAVs. The proposed method first performs an initial estimation of wireless communication quality around control terminals using a lightweight simulation based on the Longley-Rice model. Subsequently, the results of the lightweight simulation, together with data structures representing the three-dimensional terrain, are used as inputs to a machine learning model. Furthermore, field experiments are conducted in an orchard located in a mountainous area. The results confirm that the received signal strength estimated by the proposed method achieves an error within 2 dB compared to measured values.

Digital systems increasingly rely on user location data, raising significant privacy concerns. This study proposes a privacy-preserving location data utilization system that eliminates the need for dedicated base stations by integrating blockchain technology with zero-knowledge proof scheme. Our system converts data from smartphone trajectory data into zero-knowledge proof values and records only these proof values on the blockchain. Thus, the system enables verification of user movement without revealing sensitive information. By integrating the entire process with smart contracts on the blockchain, our system automates transaction processing and monetary transfers without relying on any specific organization. We conduct an experimental evaluation on the blockchain using trajectory data collected from a smartphone application.

Reservoir computing (RC) is an efficient framework for temporal signal processing in which only the readout layer is trained. Recently, physical implementations of RC have attracted increasing attention for energy-efficient and high-speed computation. In such physical systems, complex network structures are difficult to implement due to hardware constraints, even though the network topology plays a critical role in determining computational performance. Among various reservoir architectures, simple cycle reservoir (SCR) is well suited for hardware implementation due to its minimal connectivity and strong memory capability. However, its unidirectional ring structure inherently limits long-range information propagation, thereby reducing its ability to capture complex temporal dependencies. To overcome this structural limitation, we propose a reconfigurable SCR framework with a time-varying connectivity design. Selective rewiring enhances information propagation, while temporal switching among topologies increases dynamical diversity. Experimental results show that increasing the number of shortcut connections enhances second-order information processing capacity. Temporal topology switching further reduces prediction error in the NARMA10 benchmark task. These results demonstrate that the proposed approach improves computational capability while preserving hardware efficiency.

Demand for AI processing in edge computing (edge AI) has increased recently. Stochastic Computing (SC) has attracted attention as a method to reduce power consumption and circuit area, which are major constraints in edge AI. While SC offers simple arithmetic circuits, high parallelism, and excellent error tolerance, its integration with conventional binary memory requires signal conversion circuits, which offset the inherent advantages of SC. To address this issue, this study proposes a compact stochastic memory circuit that directly utilizes Random Telegraph Noise (RTN), an intrinsic noise in CMOS nanowires (NWs). The proposed circuit generates stochastic binary signals by periodically comparing and amplifying multi-level noise from the NWs using a clocked-inverter latch. Furthermore, we introduced a mechanism to modulate and store output probabilities continuously by applying an offset to the noise voltage via a series-connected variable resistor. Verification using a circuit simulator (ngspice) and statistical analysis (Python) confirmed that the proposed circuit operates as designed in terms of both stochastic signal generation and probability control. These results contribute to the realization of highly efficient stochastic in-memory computing without the need for signal conversion circuits.

This study investigated a pulse density modulation (PDM) method inspired by the vestibulo-ocular reflex (VOR) neural network for analog-to-digital conversion. The proposed method employed multiple integrate-and-fire neurons (IFNs) with device variability to collectively encode input signals whose frequencies exceeded the response limit of an individual IFN. Using a behavioral model, we evaluated signal conversion performance in terms of signal-to-noise ratio (SNR). Numerical simulations demonstrated that increasing the number of IFNs improved SNR, while device variability suppressed overlaps in firing timings and further enhanced performance, indicating the potential of the proposed architecture for signal conversion beyond the response capability of a single IFN.

This study proposes a parallel Time-to-Digital Converter (TDC) architecture for measuring high-density pulse trains arriving at nanosecond intervals with picosecond-order resolution. In a conventional single-TDC configuration, internal dead time causes event loss under high-frequency input conditions. To address this issue, multiple Vernier TDCs are arranged in parallel, and input events are dynamically assigned to available channels, enabling spatial distribution of dead-time effects. First, the proposed architecture is examined under periodic input conditions to illustrate its fundamental operation and underlying concept. Then, more realistic input conditions incorporating temporal fluctuations are considered, and the relationship between input signal characteristics and the variation in measurement accuracy with increasing circuit scale is investigated. These results provide insights into key design considerations for high-speed pulse-train measurement systems.

Bluetooth Low Energy (BLE) is widely used in small IoT devices, where further power reduction is required for long-term operation. Previous studies achieved power savings while maintaining communication reliability by adaptively selecting the transmission channel and advertising interval based on a bandit algorithm, treating all combinations as options. This paper investigates the impact of selection structure on performance by comparing this combinatorial method with two alternatives: an independent method selecting the channel and interval separately, and a hierarchical method determining the interval according to the selected channel. Experimental results show that the effectiveness of each selection structure depends on the interference condition, and that the proposed methods reduce power consumption compared with the conventional combinatorial method under the evaluated conditions.

This study investigates a spatial visualization method for indoor environmental conditions integrating a sensor network with computational fluid dynamics (CFD). Temperature and carbon dioxide concentration were estimated by integrating sparse sensor observations with CFD-based background fields through Optimal Interpolation (OI). We demonstrate that the proposed OI-based method outperforms conventional interpolation techniques (e.g., IDW, Kriging) in terms of reconstruction accuracy, particularly under sparse sensor configurations. By integrating CFD-derived background fields with observational data, our approach effectively mitigates the estimation errors caused by sensor scarcity, thereby enabling physically consistent spatial visualization.

This paper develops a new systematic control method of traffic signals for a traffic flow model represented by the burgers’ cellular automaton in order to reduce traffic jams. First, we construct ”the Burgers’ cellular automaton traffic flow model (BCATFM),” which is an ultradiscrete-type traffic flow model, and formulate an optimal traffic signal control problem for the BCATFM with a cost function defined by the total number of the traffic jam for the model as a nonlinear integer programming problem. Next, in order to solve the optimization problem, we propose an optimal switching algorithm of traffic signals based on particle swarm optimization, and derives an extended algorithm for disturbances on inflow/outflow of cars based on model predictive control. Then, some numerical simulations show that the new traffic signal control methods can reduce the total number of the traffic jam in comparison with other methods, and hence it is effective.

In this paper, we investigate the dependence of path length and complexity in pattern lock authentication, focusing on unconscious individual features during swipe motions. It has been shown that personal authentication based on unconscious individual features is both practical and robust, achieving low False Rejection Rate (FRR) and False Acceptance Rate (FAR). These results were obtained using a fixed pattern length in the pattern lock. Variations in pattern length or complexity may lead to changes in users’ unconscious behavioral characteristics and the number of extracted feature points, which could, in turn, affect authentication performance. Therefore, the purpose of the present paper is to investigate how authentication accuracy depends on path length and complexity in swipe motions. We conducted authentication experiments across three levels of path length and complexity. The results indicate that path length has only a minor effect on FRR, while path complexity has a modest impact on accuracy. More complex paths improve performance, although even simple paths maintain reasonable accuracy (average FRR: 16.4%). Therefore, the authentication method based on unconscious characteristics in swipe motions is both practical and secure for real-world smartphone applications regardless of path length or complexity.

A triangular ring network of delay-coupled excitable multivibrators with inhibitory control was investigated as a minimal physical platform for clockless spatiotemporal information processing. Each unit is a light-sensitive excitable multivibrator that exhibits hybrid dynamics composed of a continuous internal state and discrete output switching. Three such units were optically coupled in a closed loop with an explicit time delay, and localized excitatory and inhibitory stimuli were applied to examine how spatiotemporal input conditions are mapped onto collective dynamical states. The experiments revealed input-dependent selection between transient propagation and persistent rotating-wave synchronization. The rotation direction, clockwise or counterclockwise, can be controlled by either removing a coupling link or applying an inhibitory input to a specific unit, indicating deterministic state selection governed by the interplay of delay, refractory dynamics, and loop structure. A reduced circuit model reproduces the observed responses and clarifies the role of inhibitory control in reshaping propagation pathways. These results show that the proposed network can distinguish differences in the location and timing of inputs without an external clock and retain the response as an internal dynamical state.