ABSTRACT. Despite significant advancements in integrating smart technologies in industrial environments in recent years, many factories still rely on traditional paper-based methods. To bridge this technological gap and accelerate the adoption of smart solutions, we have developed the Smart E-Ink Tag System. The RSSI-based BLE Smart Tag localization feature is one of the system's core functions. However, industrial environments pose challenges such as abundant metal structures, obstacles, and widespread multipath fading, making data collection difficult due to the need for fingerprinting without disrupting ongoing manufacturing processes. In this study, we collect fingerprint data in a real-world bolt manufacturing factory and apply various data augmentation techniques as well as machine learning-based prediction methods, including WkNN, Random Forest, MLP, and LSTM. We compare the performance of these techniques to determine the most suitable combination for application to the Smart E-Ink Tag System.

ABSTRACT. Photonic decision making has been conducted to solve the multi-armed bandit problem, which is one of the important reinforcement learning problems. We investigate photonic decision making using a multimode semiconductor laser with optical feedback and injection. Decision making is conducted by controlling chaotic mode-competition dynamics via optical injection. We experimentally found that the mode concentration of the chaotic multimode laser can be enhanced for positive wavelength detuning, which provides fast decision-making performance.

ABSTRACT. Global positioning system (GPS) technology has changed many aspects of daily life. However, there is a limitation that it is difficult to estimate indoor location. Therefore, to solve this problem, ultra-wide-band (UWB) with high temporal resolution of nanoseconds was proposed. To do this, the distance must be measured based on the time it takes for the signal transmitted from the Initiator to pass through the Responder. However, due to the nature of the indoor environment, there are many obstacles, and there are difficulties in positioning due to non line-of-Sight (NLoS) rather than line-of-sight (LoS). Therefore, in this paper, the NLoS signal is detected from the received signal. suggest a method.

ABSTRACT. The relationship between music preference and phase difference (PD) between electroencephalographic and simultaneously recorded music audio was tested for statistical significance in time-frequency domains. Experimentation was conducted using unique 40-second non-vocal music excerpts, for 30 trials of listening and rating the music preference. The preference was converted into 2-class disliked (D), liked (L). EEG at six channels and music audio were simultaneously recorded. The signals were band-pass filtered, Hilbert transformed, before PD between EEG and Audio. Significant (p < 0.05) differences of EEG-Audio PD between D/L music preference in frontal regions were found β-rhythms, γ-rhythms, and δ-rhythms. The PD of EEG frequency was dependent on the individuals. It suggests a tendency for individuality. Significant PD for D/L music preference was differentially induced at the beginning and the end of the music-listening. This study could lay a foundation for music recommendation systems using novel EEG-audio PD in time-frequency domains to detect music preference.

ABSTRACT. Nonverbal vocalizations, such as coughing and sniffling, play a crucial role in the diagnosis and assessment of respiratory diseases. The prolongation of respiratory diseases can worsen health, thus accurate diagnosis and management are essential. For this, the development of an efficient and reliable healthcare system based on deep learning is necessary. The irregularity of nonverbal vocalizations poses challenges for deep learning inference, but this paper proposes the use of low-cost, low-power, and compact microcontroller units (MCUs) for user-friendly on-device systems. Although model lightweighting is essential in resource-limited MCU environments, it can lead to performance degradation. To address this, the paper suggests a Split and Combine (S&C) data augmentation technique and a Residual Depthwise Separable Convolution (ResDSC) deep learning model, considering the characteristics of nonverbal vocalizations. It maintains robust performance through noise injection and volume adjustment techniques. The proposed model is optimized using knowledge distillation and quantization techniques and tested for on-device performance on ES32 and Raspberry Pi Zero 2 W.

ABSTRACT. Since GPS (Global Positioning System) cannot meet the accuracy requirements indoors, indoor Location-Based Services (LBS) have become increasingly important. BLE (Bluetooth Low Energy) offers cost and accuracy advantages. Typically, the position fingerprinting method is used for indoor positioning. However, the indoor environment introduces various offsets in Bluetooth RSSI (Received Signal Strength Indicator). This study proposes a method to correct RSSI offset using SVR (Support Vector Regression), supplemented with particle filter post-processing to reduce the positioning error. Experiments performed by k-NN (k-Nearest Neighbors), BPNN (Back-Propagation Neural Network), and SVR-based position fingerprinting methods, results show that the proposed method can effectively reduce the positioning error.

ABSTRACT. In today's general hospitals, the workload for nurses can be overwhelming, with one nurse often responsible for as many as 16 or more patients. This situation is exacerbated in closed wards, where patients require both physical and mental care, yet nurses still find themselves caring for the same high number of patients as in regular wards. Mental patients, in particular, require close monitoring due to their unpredictable actions, but this becomes challenging due to staffing shortages. To address these challenges, we propose the implementation of an automatic abnormal action recognition system. This system utilizes real-time video monitoring to detect and track the locations of mental patients and identify their abnormal actions. To enhance accuracy, we have defined specific abnormal actions that typically occur in closed wards and have developed a rule-based algorithm to identify them. The entire process is implemented using the DeepStream SDK for high-speed performance, with an impressive processing time of approximately 0.01 seconds. The recognition accuracy of the system averages over 90\%. The improved system is expected to provide significant benefits to healthcare providers. It will enable them to make more informed judgments about the conditions of mental patients, reducing the likelihood of dangerous situations and improving the overall working environment for healthcare professionals. This not only promotes the well-being of patients but also enhances the job satisfaction and safety of healthcare workers.

ABSTRACT. The demand for connections in Internet of Things (IoT) networks has been increasing significantly. Accordingly, network providers are facing many challenges in order to provide services to users. One of the important challenges is to manage and allocate the radio resources, including time, frequency, and energy resources. In this paper, we propose to utilize graph neural networks (GNNs) to develop scalable methods for solving the energy efficiency optimization problem in IoT Networks. First, we convert the system into graph-structured data. Then, we propose an edge convolution layer that learns the optimal user association in the unsupervised scheme. Various simulations will evaluate the performance of the proposed edge convolution layer compared to the existing method.

ABSTRACT. Elementary cellular automata (ECAs [1]) are simple digital dynamical systems where time, space, and state variables are all discrete. Depending on simple rules of simple Boolean functions, the ECAs generate various binary periodic orbits (BPOs). Real/potential applications of the BPOs include error correcting codes [2], time-series approximation [3] and control signal of walking robots [4]. The ECAs are well suited for FPGA based hardware implementation [5]. This paper introduces mixed-rule cellular automata (MCAs) governed by two rules (rule pair) of simple Boolean functions. Depending on the rule pair, the MCAs can generate various BPOs with strong stability. As a first step to engineering applications, we consider typical BPOs corresponding to control signal of switching power converters. Performing precise numerical experiments, we have discovered suitable rule pairs in which the MCA generate BPOs with strong stability. Such strong stability is impossible in ECAs. Presenting simple FPGA based hardware prototype, the typical BPOs are confirmed experimentally.

[1] S. Wolfram, Cellular automata and complexity: collected papers, CRC Press, 2018 [2] D. R. Chowdhury, S. Basu, S. I. Gupta, P. P. Chaudhuri, Design of CAECC - cellular automata based error correcting code, IEEE Trans. Comput., 43, pp. 759--764, 1994. [3] O. Yilmaz, Symbolic computation using cellular automata-based hyperdimensional computing, Neural Computation, 27, pp. 2661-2692, 2015. [4] T. Suzuki and T. Saito, Synthesis of three-layer dynamic binary neural networks for control of hexapod walking robots, Proc. IEEE/CNNA, 2021. [5] T. Okano and T. Saito, Permutation elementary cellular automata: analysis and application of simple examples, M. Tanveer et al. (Eds.): ICONIP 2022, LNCS 13623, pp. 321-330, 2023.

ABSTRACT. Out-of-distribution (OOD) data detection is an unavoidable safety issue when looking at real-world applications of AI. In this study, we focus on BNNs, which can estimate the uncertainty inherent in AI inference on a cause-by-cause basis, and evaluate the basic out-of-distribution data detection performance of BNNs. In addition, we will propose and evaluate data preprocessing to improve the detection performance of out-of-distribution data for adaptation to more advanced problems. As a result, we confirmed that the detection performance is increased by using high-frequency components extracted by the wavelet transform, especially the 5/3 lifting wavelet transform, for learning.

ABSTRACT. Image steganography is a technique for embedding secret messages in images. SteganoGAN, one of the previous methods, uses generative adversarial networks and achieves high payload when the input is a noise-free image. However, real images contain real image noise (RIN) generated during the image acquisition process, which can degrade the performance of SteganoGAN. We propose a RIN-aware image steganography that uses a real image denoising method as preprocessing and modifies the loss function of SteganoGAN. These modifications encourage the proposed method to embed a secret message that simulates RIN into a pseudo-noise-free image obtained by denoising. Experimental results show that the proposed method improves the trade-off between image quality and payload compared to the previous method.

ABSTRACT. This study applied the machine learning methods for sentiment analysis and topic classification of news titles. Based on diverse news dataset, sentiments of news were categorized as 'Positive', 'Negative', and 'Neutral'. The topics were identified using a Support Vector Machine (SVM) classifier with Term Frequency-Inverse Document Frequency (TF-IDF) vectorization. The SVM classifier, paired with TF-IDF vectorization, exhibited stellar efficacy, achieving an overall accuracy rate of 97.85%. It was found that notable sentiment distributions among news sources and offer insights into the biases and patterns of modern news reporting, demonstrating the efficacy of machine learning in media analysis.

Presentation type: Poster

ABSTRACT. This paper delves into network caching optimization within the 5G context, addressing the limitations posed by the 'black box' nature of AI-driven strategies. Centered around eMBB, URLLC, and mMTC services, it introduces mathematical models for allocation and replacement strategies, aimed at enhancing transparency and efficiency in caching processes. The objective of this study is to identify more effective strategies for network caching in the evolving 5G network environment.

ABSTRACT. In this work, we introduce a machine learning model combined from Graph Neural Network (GNN) and Federated Learning framework (FL). We evaluate the proposed model on power control problem in small-cell wireless system, which has different number of users in each cell and the proposed model is expected to maximize the sum rate of whole system. The proposed model shows the more effective performance on maximize total data-rate task and have more adaptability compare to benchmarks.

ABSTRACT. Earth-Moon-Earth (EME) communication is a technology that involves transmitting signals from Earth to the Moon and receiving signals reflected off the lunar surface. When the reflected signals reach Earth, they experience approximately 90% loss, necessitating the reception of ultra-low Signal-to-Noise Ratio (SNR) signals. As the SNR decreases, the probability of detecting the signal itself diminishes, underscoring the importance of efficiently detecting the preamble, which represents the initial part of the signal. Therefore, this paper proposes a method for Preamble Detection in an ultra-low SNR environment, specifically tailored for long-distance communication, utilizing LoRa (Long Range) signals and spectrogram images representing the energy of the signal. Deep Neural Network (DNN) is employed to perform effective Preamble Detection even in environments with extremely low SNR.

ABSTRACT. We classified hateful comments using BERT embedding techniques such as KcELECTRA and KcBERT. The trained the model for the dataset of Korean hate expressions. When comparing both KcBERT and KcELECTRA pretrained using Korean comment data, the classification model using KcELECTRA showed better performance. This study will contribute to determine what percentage of comments on a specific article are affected by malicious comments, and it is also possible to black box hate comments.

ABSTRACT. In this paper, we introduce a way to improve the distributed learning performance at the server via gradient analyzing without the need of data on server. Therefore, the FL can achieve a significant improvement in performance while still guarantee communication efficiency.

ABSTRACT. The indoor wireless positioning system based on Ultra-Wideband (UWB) technology has limitations in accuracy, particularly in non-line-of-sight (NLOS) environments. Therefore, distinguishing between line-of-sight (LOS) and NLOS channels is an important research challenge. In this paper, we propose using a deep learning model that combines a Transformer and Convolutional Neural Network (CNN) for the classification of NLOS channels based on the channel impulse response (CIR) of received UWB signals. We conducted experiments using an open-source dataset and demonstrated that the CNN-Transformer model outperforms benchmark algorithms in channel classification performance.

ABSTRACT. Color correction of dark images is not easy task because of their low contrast, so various methods have been studied. This study proposes a method of color correction method for dark images using machine learning.

ABSTRACT. Stains can be classified into water-soluble stains, oil-soluble stains and insoluble stains. To remove stains, different methods have to be used depending on the type of stains, but it is difficult for people without knowledge about stains to identify the type of stains by visual inspection. In recent years, image classification has been studied in various fields. However, there are few studies that identify the type of stains on fabric. This study aims to estimate the type of stain from an image of a stain on fabric with high accuracy using Convolutional Neural Network (CNN). For this purpose, this study makes a dataset and considers good CNN models.

ABSTRACT. The integration of the Internet of Things (IoT) and Blockchain, known as IoT-Blockchain, promises enhanced security and privacy in applications such as smart homes. However, given that many IoT devices are resource-constrained and battery-driven, it's pivotal to understand the energy implications of incorporating Blockchain technology. In this study, we design and evaluate a private IoT-Blockchain system using Ethereum on Raspberry Pi 4. We compare the power consumption of two consensus algorithms: Proof of Work (PoW) and Proof of Authority (PoA). We found that PoW consumes approximately double the power of PoA.

ABSTRACT. It is envisioned that the Internet of Things (IoT) will be strengthened with blockchain technology in the so-called IoT blockchain system. The IoT Blockchain system may abstract the heterogeneous IoT protocols and devices and provide blockchain benefits. Integrating blockchain with the IoT may ensure IoT data security, traceability, and integrity. This study investigates the integration of IoT with Ethereum Blockchain with private blockchain deployments. Moreover, we focus on latency performance in the IoT blockchain system. More specifically, we deploy various overlay networks in an Ethereum-based IoT system and evaluate their impact on latency, considering the Transaction-oriented Latency (TOL) and block-oriented latency (BOL) metrics. Our findings reveal that a fully connected graph does not lead to low latency, and the Barabási–Albert (BA) model effectively mitigates latency increases due to higher transaction frequencies. This study contributes to understanding latency in private IoT Blochain with multiple nodes, emphasizing the importance of overlay network design for optimal system performance.

ABSTRACT. In recent years，the incidence of soft errors in VLSI has been increasing due to miniaturization，higher integration，and lower supply voltages．Soft errors are temporary errors caused by the impact of high-energy particles such as alpha particles and neutrons on circuits．Recent studies have proposed latches that are tolerant to soft errors occurring at multiple nodes as VLSI scaling continues．In this paper，we propose a soft-error tolerant circuit using the DICE structure，which is a soft-error tolerant structure，for non-volatile flip-flops that have been studied to reduce the power consumption of battery-powered devices.

ABSTRACT. As the Internet of Things (IoT) advances, Wireless Sensor Networks (WSNs) are attracting attention as an underlying technology for the IoT. As a standard for WSNs, IEEE802.15.4 defines multiple channels as an option, but there needs to be more discussion on their allocation and scheduling methods. In this study, we propose an access control protocol that achieves a high packet delivery ratio and low latency by scheduling suitable for channel assignment in tree-based WSNs. We derive the number of channels required for a tree WSN and achieve a high packet delivery ratio and low latency by avoiding packet collisions through multi-channel and scheduling.

ABSTRACT. Random access protocols such as slotted ALOHA are considered suitable for mMTC, and the application of NOMA to slotted ALOHA has been considered to improve performance. However, ALOHA has a near-far problem in which devices near the base station are more likely to transmit successfully than those far away, and it is not still clear how much NOMA can mitigate the near-far problem. In this paper, we evaluate the fairness of successful transmission in different power control schemes by simulation.