Nonlinear Theory and Its Applications, IEICE
Online ISSN : 2185-4106
ISSN-L : 2185-4106
Current issue
Displaying 1-8 of 8 articles from this issue
Special Section on Resolving Nonlinear Problems by Python
  • Tetsushi Ueta
    2025 Volume 16 Issue 2 Pages 208
    Published: 2025
    Released on J-STAGE: April 01, 2025
    JOURNAL OPEN ACCESS
    Download PDF (89K)
  • Yuu Miino
    2025 Volume 16 Issue 2 Pages 209-221
    Published: 2025
    Released on J-STAGE: April 01, 2025
    JOURNAL OPEN ACCESS

    Bifurcation analysis is a critical tool for understanding the behavior of nonlinear dynamical systems. In this study, we explore the impact of tolerance settings in the RK45 method on the accuracy of bifurcation analysis. While tighter tolerances can improve the accuracy of numerical solutions, they also significantly increase the computational cost, which is particularly important in bifurcation analysis that requires numerous iterations to explore parameter spaces. Using the Duffing equation as a case study, we investigate how variations in tolerance settings affect the observed bifurcation phenomena. We conduct numerical experiments to compare the performance of the RK45 method with that of the 4th-order Runge-Kutta method (RK4), focusing on the accuracy of the solution trajectories and the computational efficiency. Our results show that default tolerance settings in the RK45 method may fail to capture certain bifurcation phenomena, potentially leading to misleading conclusions. Interestingly, the bifurcation diagrams obtained by varying rtol revealed a rich variety of bifurcations, representing a novel discovery not previously reported. To address this, we identify a range of tolerance values, approximately 1×10-4 to 1×10-5, that balance accuracy with computational efficiency. These findings offer practical guidelines for selecting appropriate tolerance settings in numerical integration, ensuring accurate bifurcation analysis while minimizing computational resources.

    Download PDF (4551K)
  • Anna Wróbel, Yulia Sandamirskaya, Thomas Ott
    2025 Volume 16 Issue 2 Pages 222-232
    Published: 2025
    Released on J-STAGE: April 01, 2025
    JOURNAL OPEN ACCESS

    Near-infrared (NIR) spectroscopy is widely used in agriculture and the food industry to classify fruits and determine ripeness, soluble solids content, pH and acidity. Neuromorphic technology offers the potential for low-power real-time analysis systems based on NIR spectroscopy signals. This study presents a development pipeline for a neuromorphic classifier using Spiking Neural Networks (SNNs) to classify NIR spectra of fruit species. The SNN-based algorithm is implemented in the DYNAP-SE neuromorphic device. The classifier's performance is compared to non-spiking Artificial Neural Networks (ANN) and Support Vector Machines (SVM). Python is used throughout the development, showcasing its versatility as a development tool.

    Download PDF (2939K)
  • Masato Izumi, Kenya Jin'no
    2025 Volume 16 Issue 2 Pages 233-249
    Published: 2025
    Released on J-STAGE: April 01, 2025
    JOURNAL OPEN ACCESS

    In recent years, advancements in artificial intelligence, especially in natural language processing (NLP) models, have progressed rapidly. These models demonstrate remarkable results through training on large datasets and extensive architectures. However, the output process is often a black box, and the decision-making process remains unclear. Our research focuses on the internal representations, specifically the latent variables, generated by NLP models. In earlier work, we explored the latent variables and spaces produced by Sentence-BERT using image generation models. This approach aimed to visualize these spaces by converting discrete textual embeddings into images, introducing continuity and revealing novel relationships. This paper presents the development of an image generation model that uses a common decoder for both GPT-2 and Sentence-BERT, aimed at examining how differences in model architecture affect their latent spaces. We also investigate the impact of training dataset differences by comparing models trained in English and Japanese. Our findings indicate that while the models often generate similar outputs, significant differences emerge in sentences containing multiple elements, attributable to the differing focuses and objectives of the models. Our goal is to understand these latent spaces and contribute to the development of explainable AI.

    Download PDF (9031K)
  • Kyosuke Kageyama, Takeshi Kumaki
    2025 Volume 16 Issue 2 Pages 250-270
    Published: 2025
    Released on J-STAGE: April 01, 2025
    JOURNAL OPEN ACCESS

    Recently, the digital images are used everywhere. The digital images are need at an investigation, court case, our daily life, and so on. However, the digital images are easily edited by anyone. Therefore, the digital image is required to have authenticity. The morphological pattern spectrum has been proposed as a novel technology to detect manipulation. This proposed technology was used a total pixel values approach, which can detect manipulation from the change of the pixel values in an image. This approach can detect the manipulated image and also judge the rotated image without manipulation from the original image. Therefore, this approach can detect manipulation with high accuracy. However, this approach has a problem not to judge a compressed image like JPEG format because the compressed image is changed the pixel values from the original image. The morphological pattern spectrum used a pixel-scale counting approach is proposed as novel technology to improve this problem. This improved approach is counting the number of same pixel scale as a structuring element size. It is implemented in Python for the purpose of being embedded on mobile devices and using with AI technology in the future. Therefore, this improved approach can detect the compressed image because it isn't affected by the change of the brightness values. In this paper, this improved approach is verified to judge the original image and the JPEG image with and without manipulation from the original image. In addition, the pixel-scale counting-based morphological pattern spectrum and the famous existing technology are compared. From these results, the pixel-scale counting approach based morphological pattern spectrum can detect manipulation from the JPEG image, and is confirmed superiority in detection accuracy, detection capability understandable for human eye, supports a variety of image formats, and usability than the existing technology.

    Download PDF (9289K)
Regular Section
  • Takafumi Kunimi, Kota Ando, Takao Marukame, Tetsuya Asai
    2025 Volume 16 Issue 2 Pages 271-289
    Published: 2025
    Released on J-STAGE: April 01, 2025
    JOURNAL OPEN ACCESS

    The free-energy principle (FEP) is a new theory proposed in the field of theoretical neuroscience to mimic the processing of brain information. This FEP posits that the brain functions to minimize a single cost function called variational free energy, suggesting a relationship with existing brain theories, such as predictive coding (PC). By leveraging this principle, a top-down approach could be adopted to mimic the overall information processing in the brain, in contrast to the bottom-up approach of traditional neuromorphic computing components. In this study, we designed a PC network based on this principle using classical analog circuits and evaluated its performance using the circuit simulator NGSPICE. Analog circuits are more energy-efficient, integrate better, and more accurately mimic biological information processing than digital circuits. Consequently, the circuit demonstrated a performance similar to that of simulations conducted in Python, indicating the potential for creating novel brain-inspired hardware.

    Download PDF (262008K)
  • Hiroki Shinagawa, Gouhei Tanaka
    2025 Volume 16 Issue 2 Pages 290-307
    Published: 2025
    Released on J-STAGE: April 01, 2025
    JOURNAL OPEN ACCESS

    Spiking neural networks (SNNs) are the basis of low-power neuromorphic computing systems and hardware. Most previous studies on unsupervised SNNs have tested their ability in grayscale image classification. In this study, we mainly propose color opponency-based filters, inspired by retinal color vision, for color image classification with an unsupervised two-layer SNN and a simple linear classifier. We demonstrate that the color opponency-based filters in multiple visual pathways are more effective than conventional grayscale transform methods in a color image classification task with ETH-80 dataset. Our results suggest that biological color vision mechanisms can expand the potential of shallow SNN models.

    Download PDF (2343K)
Errata
feedback
Top