This paper proposes a hardware architecture for edge AI applications of Bayesian neural networks, enabling uncertainty evaluation in neural network inference outputs. A lightweight and highly efficient architecture originally designed for binary neural networks is extended to support Bayesian inference by introducing Monte Carlo sampling based on the Bernoulli distribution instead of the Gaussian distribution, thereby eliminating multiplication by restricting weights to binary values (-1/+1). The random number generators required for Monte Carlo sampling are also significantly downsized by sharing bitstreams. The FPGA prototype demonstrates a high simulated efficiency of 63.08 GOPS/W at 50 MHz.

Wireless Brain-Inspired Computing (WiBIC) integrates spiking neural network (SNN) functionality directly into IoT nodes, unifying the concepts of networking in both the IoT and neural domains. Previous studies have shown that WiBIC enables intelligent information processing within the network itself. To extend the applicability of WiBIC, it is essential to show that the network can autonomously generate biologically inspired responses. This paper presents a study on modeling Pavlovian conditioning using the WiBIC platform. We use the Modulated Spike-Timing-Dependent Plasticity (MSTDP) learning rule to facilitate associative learning. For wireless communication within the WiBIC network, we use Asynchronous Pulse Code Multiple Access (APCMA), which effectively emulates the behavior of biological networks. The WiBIC platform was implemented on a Field Programmable Gate Array(FPGA). Experimental results confirm that the implemented system successfully replicates the Pavlovian conditioning process, demonstrating the ability of WiBIC to emulate biological learning mechanisms.

In this study, we propose a new semi-supervised learning (SSL) method that achieves accuracy comparable to FixMatch with a smaller batch size. Our method generates multiple strong augmentations from a single unlabeled data point and applies Mixup regularization to enhance training stability. We also prioritize effective data augmentation algorithms. We evaluated our method by comparing its accuracy and computation time to FixMatch, finding that generating five strong augmentations from a single unlabeled data point provided the highest accuracy of 94.13% and reduced computation time to 70.8% of FixMatch. Additionally, our method outperformed other SSL methods on CIFAR-10, CIFAR-100, SVHN, and STL-10, especially with fewer labeled samples. An ablation study confirmed that both Mixup regularization and Prioritized Strong Augmentations contribute to improved accuracy and stability. Our method thus achieves comparable accuracy to FixMatch while reducing computation time.

We propose a prediction method based on reservoir computing for capturing transitions between local attractors in time series of intermittent chaotic systems. We investigate the prediction performance focusing on whether the predicted signal accurately captures the timing of transitions. The results in this paper show that the prediction signal can successfully predict the transition behavior of the target signal in single-step-ahead predictions, while increasing the prediction steps leads to delayed or failed transitions. To address this, we introduce a criterion of permissible delay step, which permits for small timing errors when prediction precision is evaluated. This approach significantly improves the transition prediction precision, especially in cases with intermediate prediction steps. Our findings provide insights into the potential of reservoir computing for predicting critical transition events in nonlinear dynamical systems.

In this study, we propose a method for acquiring dynamical models of machines performing repetitive motions using Reservoir Computing (RC), with a view toward future applications in anomaly detection for real-world systems. Unlike conventional approaches that utilize random noise as input, the proposed method employs cyclic input signals, enabling safe and stable training of RC under real machine conditions. In this framework, RC is trained on output time-series data under normal operating conditions, and anomalies are assumed to be detected based on the discrepancy between the predicted and observed outputs. This RC configuration based on cyclic inputs has the potential to serve as a core component in anomaly detection systems for future real-world implementations. To investigate the effect of input diversity, five levels of cyclic input patterns with varying degrees of randomness were defined. Furthermore, two methods were employed to determine the parameter randomness. Simulation results demonstrated that even simple generated randomness maintains the prediction accuracy, confirming that RC can successfully acquire the dynamical behavior of machines exhibiting cyclic motions within a reduced complexity.

Image denoising aims to remove noise that is inevitably generated in the image acquisition process, and is still a fundamental problem in image processing. In particular, there is a lot of interest in removing real image noise that occurs in actual devices, instead of synthetic noise such as Gaussian noise. Existing real image denoising methods that use convolutional neural networks (CNNs) or transformers tend to have extremely large parameter sizes and computational costs, which makes it impossible to use these methods on smartphones and embedded devices. Gated texture convolutional neural network (GTCNN), proposed by Imai et al., achieves a better trade-off between the number of parameters and performance on synthetic noise by using two independent CNNs for context extraction and denoising based on that, and combining the results using a gating mechanism. We propose a new variant of GTCNN that further improves the trade-off between the number of parameters and performance of the method, i.e., parameter efficiency, by replacing specific convolutional layers of GTCNN with depthwise separable convolution, and investigate its performance on real image noise. Since the selection of layers to which the replacement is applied is not obvious, we identified the optimal replacement points that do not impair performance through comprehensive experiments. The experimental results on existing real image noise datasets showed that the proposed method achieves significant improvements in parameter efficiency compared to existing image denoising methods, including lightweight methods such as GTCNN.

This paper analyzes Convolutional Neural Networks (CNNs) from a spatial frequency perspective using Z-transforms to evaluate kernel transfer functions. By examining which frequency bands kernels respond to, we quantitatively analyze feature processing through CNN layers. Experiments with custom and pre-trained models (InceptionV3, VGG16, ResNet-50, DenseNet-121) reveal CNNs capture high-frequency features in early layers while increasingly focusing on low-frequency features in deeper layers. Kernel pruning experiments demonstrate that kernels with larger weights capturing low-frequency features are critical for model performance. These findings provide insights into CNN feature extraction processes, contributing to improved interpretability and model construction guidelines.

Many zero-shot image restoration methods have been proposed by leveraging pre-trained image diffusion models. These methods are capable of performing various image restoration tasks without the need for task-specific training. In general, such methods tend to improve restoration performance by using conditional image diffusion models, such as those based on classes. However, the challenge has been that a separate method is required to determine the appropriate class from degraded images. In this study, we focus on image colorization and propose a method in which the classification of the input grayscale image is performed by applying CLIP, a type of vision-language model, and the resulting class is used as the class condition for a conditional image diffusion models. Through experiments, it was confirmed that the proposed method enables high-precision colorization compared to conventional zero-shot image restoration methods without class conditions, as well as deep neural networks specifically trained for image colorization.

Prompt engineering is the technique of designing and adjusting prompts for Large Language Models (LLMs) and generative AI, such as ChatGPT, to achieve the desired output. By leveraging this method, it is possible to maximize the AI's knowledge and capabilities, enabling the efficient generation of high-precision text. The importance of prompt engineering lies in the natural language processing characteristics of AI. LLMs learn statistical patterns from vast amounts of text data and generate outputs based on the given inputs. However, since the results are highly dependent on how prompts are formulated, effective prompt design, along with its evaluation and refinement, is essential for achieving the desired outcomes. In this study, we utilize prompt engineering to automate the generation of conversational data for Natural Language Processing (NLP) tasks. Specifically, we focus on generating dummy data for training and evaluating machine learning models, as well as for developing dialogue systems. Furthermore, we vectorize the generated conversational texts and apply dimensionality reduction techniques for visualization, allowing us to analyze the diversity of the conversations and identify clustering tendencies. This approach helps us examine how different prompt designs influence the outputs and reveal the distribution characteristics of the generated data. Through this analysis, we can assess the suitability of the generated data as dummy data for NLP applications.

This study proposes an AI-driven learning behavior analysis and modeling framework aimed at supporting adaptive learning in English education. By incorporating multimedia technology, the system combines machine learning models with Internet of Things (IoT) technology to analyze students' behavioral performance online during the learning process and evaluate their learning outcomes. The core methodology of this framework includes the following steps: First, data preprocessing is conducted to optimize feature representation, and K-means clustering, along with Principal Component Analysis (PCA), is used to identify student behavior patterns and extract key features. Next, based on the clustering results, Mean Impact Value (MIV) analysis is applied to quantify feature importance and eliminate redundant information. Finally, multiple AI models, such as Bidirectional Long Short-Term Memory (BiLSTM), Temporal Convolutional Network (TCN), and Random Forest (RF), are integrated to capture temporal dependencies and spatial behavior patterns in learning tasks. We validated this method in four core English learning tasks: listening, speaking, reading, and writing. Experimental results show that the feature extraction strategy based on K-means-PCA-MIV optimization significantly improves the model's performance in complex evaluation scenarios. Among all the evaluation tasks, the Random Forest (RF) model performed the best, while BiLSTM demonstrated strong capability in modeling temporal data. Additionally, we found that writing tasks, due to their structured nature, are easier to assess, whereas speaking tasks present more challenges due to individual variability. This study provides new insights into intelligent assessment and personalized learning path design in English education and supports behavior data-driven dynamic learning intervention strategies, promoting the deep integration of complex communication science and multimedia technology in the education field.

In recent years, Information-centric networking (ICN) has attracted wide attention as a new network architecture which can efficiently deliver digital content and Internet of Things (IoT) data. In ICN, we can request data by content name and send request packets by referring to FIB (forwarding information base) of routers. However, unlike IP communication, the ICN router's FIB does not depend on location, so prefix aggregation is difficult. On the other hand, CDN (content delivery network) is widely used as a technology to improve the delivery quality of users and reduce the amount of traffic in networks. Using CDN, it can replace the location of distribution of original content provided by the publisher. In this paper, we propose to use the CDN as a method of allocating originals of content and replace them to reduce the FIB size when using the LPM (longest prefix matching) to aggregate FIBs. When replacing originals of contents, we also consider reducing the network load as well as reducing the FIB size. Through numerical evaluation using multiple types of topologies, we clarify the influence of evaluation weights of topologies on the FIB size and the network load.

Spectral graph theory has been an active research area, analyzing the structural properties of networks using the eigenvalues and eigenvectors of matrices, e.g., adjacency matrix and Laplacian matrix, that represent their structure. In this paper, we focus on a network representation called the spectral path, proposed by Jin et al. The spectral path is defined as a trajectory connecting low-order spectral moments — namely, the second-, third-, and fourth spectral moments — of subgraphs with different sizes. The spectral path is thought of as an embedding, and it can be useful for applications such as network classification. The aim of this paper is to explore the potential of the spectral path. More specifically, we try to answer the following research question: How does the sampling strategy affect the spectral path? Through experiments, we examine the effect of sampling strategies and demonstrate the effectiveness of combining multiple sampling strategies.

In this work, we present a novel interrupted circuit model with guaranteed chaos, and we investigate the stabilization effect of a periodic threshold on the circuit operation. Previously investigated systems exhibit stable operation states for some range of parameter values, which is not true of our model. We obtain an exact solution for the circuit equation and derive the associated return maps. Furthermore, we define the conditions that induce bifurcations in the circuit, and we obtain the bifurcation curves describing the system. Finally, we demonstrate that the use of periodic thresholds stabilizes the operation of the circuit considered here.

Vital signs are important indicators for assessing an individual's health status, and technologies that measure the vital signs of patients and caregivers using electronic devices and monitor their behavior are gaining increasing attention. Texas Instruments Inc. has developed a noncontact vital sign measurement algorithm that utilizes millimeter-wave (mmWave) radar technology. This algorithm is capable of detecting clearer heartbeat signals using an infinite impulse response (IIR) filter to reduce and eliminate interference from respiration. However, the heart rate component in biological signals is inherently weak and susceptible to noise interference during measurement, necessitating advanced technical solutions to improve the measurement accuracy. This study proposes a novel method to enhance measurement accuracy by integrating the directional and temporal characteristics of chest movements resulting from respiration and heartbeats rather than suppressing or removing the influence of respiration as in the conventional method. The experimental results show that the error rate of the measurement results obtained by the proposed method is 0.9% at its lowest, achieving a reduction of approximately 40% compared with the error rate obtained by the conventional method.

In this study, we propose a method to construct an image recognition model with a single-channel input using transfer learning and data augmentation for music emotion classification. The data augmentation method generates a variety of spectrogram images by varying the STFT window size in small increments. This method ensures data equivalent to five times the amount of the original data and prevents degradation of classification performance due to insufficient data. The model construction method using transfer learning for grayscale images is designed to adapt the pre-trained EfficientNetV2 model, which was originally trained on ImageNet. The constructed model through transfer learning and our proposed data augmentation method achieved a classification accuracy of 94.8% on the 4Q Audio Emotion Dataset. Thus, our construction method using transfer learning for grayscale images, combined with the proposed data augmentation method, is effective in achieving music high-accuracy emotion classification.

In solving the traveling salesman problem (TSP), the grey wolf optimizer (GWO) has demonstrated superior performance; however, it faces challenges in terms of global solution search and computational efficiency, especially as the number of cities increases. This study presents an innovative approach that integrates an adaptive large neighborhood search into GWO. This methodology is designed to expand the solution search space while reducing computational time. Through numerical experiments, we demonstrated that the proposed method significantly reduces computational time and yields superior solutions, particularly for problems involving from 400 to 6,000 cities. These results suggest a significant improvement in the solution of large-scale TSPs, providing more efficient and effective optimization tools for routing problems.

Preventing packet congestion in communication networks is a critical challenge. A conventional routing strategy based on node degree to prevent congestion has demonstrated good performance in improving the traffic capacity within communication networks. However, this approach fails to differentiate between nodes with the same degree. Furthermore, since multiple nodes with the same degree exist in complex networks, the evaluation of importance based on node degree has its limitations. Addressing this limitation, we propose a probabilistic routing strategy that utilizes gravitational centrality. Our proposed strategy, however, differentiates the importance of nodes by employing a probabilistic function based on gravitational centrality rather than node degree, thereby allowing nodes with identical degrees to have varying levels of importance. Numerical experiments demonstrated that the proposed strategy achieves more than 19% larger traffic capacity compared to the efficient routing strategy, more than 13.4% larger traffic capacity compared to the effective gravitation path routing strategy, and over 16% larger traffic capacity compared to the probability routing strategy.