Masahiro Iwahashi

The analysis of speech production based on physical models of the vocal folds and vocal tract is essential for studies on vocal-fold behavior and linguistic research. This paper proposes a speech production analysis method using physics-informed neural networks (PINNs). The networks are trained directly on the governing equations of vocal-fold vibration and vocal-tract acoustics. Vocal-fold collisions introduce nondifferentiability and vanishing gradients, challenging phenomena for PINNs. We demonstrate, however, that introducing a differentiable approximation function enables the analysis of vocal-fold vibrations within the PINN framework. The period of self-excited vocal-fold vibration is generally unknown. We show that by treating the period as a learnable network parameter, a periodic solution can be obtained. Furthermore, by implementing the coupling between glottal flow and vocal-tract acoustics as a hard constraint, glottis-tract interaction is achieved without additional loss terms. We confirmed the method's validity through forward and inverse analyses, demonstrating that the glottal flow rate, vocal-fold vibratory state, and subglottal pressure can be simultaneously estimated from speech signals. Notably, the same network architecture can be applied to both forward and inverse analyses, highlighting the versatility of this approach. The proposed method inherits the advantages of PINNs, including mesh-free computation and the natural incorporation of nonlinearities, and thus holds promise for a wide range of applications.

We propose an explainable topic modeling method that tracks user interests to elucidate their association with social events while ensuring high reliability and low computational cost. Conventional methods capture changes in topic distribution over time but cannot distinguish whether these changes are related to social events or unrelated temporal variations. Additionally, recent advances in large language models (LLMs) have enhanced topic explanation, but they remain computationally intensive and prone to hallucination. Thus, our method introduces graphical lasso-guided iterative principal component analysis (GLIPCA) and retrieval augmentation generation (RAG) into a topic modeling framework. We regard keyword frequencies from Google Trends as a time series. First, GLIPCA separates event-related trends from unrelated temporal variations by clustering time series with similar waveforms and removing indirect correlations from unrelated major trends. Second, RAG reduces the hallucination effect and computational cost by restricting topic label generation to the most relevant documents. Additionally, we use LLMs to analyze the relationship between time series and words, which were rarely addressed in previous studies. We conducted experiments using a Vietnamese newspaper and Google Trends data to study the Tet holiday and APEC in Vietnam in 2017. The results show that our proposed method effectively extracted topics before, during, and after event. These findings are consistent with the Vietnamese socio-cultural context (e.g., focus on shopping before Tet and the return to work after Tet). This demonstrates that our method explored data-driven insights that allowed for systematic and quantitative descriptions of changes in user interests rather than only qualitative observations.

The analysis of speech production based on physical models of the vocal folds and vocal tract is essential for studies on vocal-fold behavior and linguistic research. This paper proposes a speech production analysis method using physics-informed neural networks (PINNs). The networks are trained directly on the governing equations of vocal-fold vibration and vocal-tract acoustics. Vocal-fold collisions introduce nondifferentiability and vanishing gradients, challenging phenomena for PINNs. We demonstrate, however, that introducing a differentiable approximation function enables the analysis of vocal-fold vibrations within the PINN framework. The period of self-excited vocal-fold vibration is generally unknown. We show that by treating the period as a learnable network parameter, a periodic solution can be obtained. Furthermore, by implementing the coupling between glottal flow and vocal-tract acoustics as a hard constraint, glottis-tract interaction is achieved without additional loss terms. We confirmed the method's validity through forward and inverse analyses, demonstrating that the glottal flow rate, vocal-fold vibratory state, and subglottal pressure can be simultaneously estimated from speech signals. Notably, the same network architecture can be applied to both forward and inverse analyses, highlighting the versatility of this approach. The proposed method inherits the advantages of PINNs, including mesh-free computation and the natural incorporation of nonlinearities, and thus holds promise for a wide range of applications.

Non-invasive methods for observing the morphology of living oleaginous yeast are ideal for optimizing the production of various oils, such as food oils, ol

Water purification treatments often use the jar test to determine the appropriate coagulant dosage. In the jar test, it is necessary to develop a method for predicting future video frames after dosing a water sample with the coagulant. For this purpose, the present paper describes a multi-resolution echo state network (ESN) with two novel features. This study is the first to realize video prediction for the jar test in water purification. The first novelty of the proposed method is long-term video prediction from a small number of training videos. This overcomes the need for many training videos when applying video prediction methods developed in other fields to our task. The second novelty is a multi-resolution analysis framework for ESN-based video prediction. This solves the problem of existing recursive prediction strategies accumulating errors over time, eventually producing trivial results. The strength of our proposed method is that prediction results at multiple resolutions effectively remove errors at specific resolutions. Experiments using kaolin and zeolite show that the proposed method achieves improved accuracy compared with state-of-the-art deep learning-based video prediction methods. Furthermore, compared with conventional methods, our method provides a better quantification of the visual information contained within the jar test results after dosing water with the coagulant.

In the domain of road inspection and structural health monitoring, precise crack identification and segmentation are essential for structural safety and disaster prediction. Traditional image processing technologies encounter difficulties in detecting cracks due to their morphological diversity and complex background noise. This results in low detection accuracy and poor generalization. To overcome these challenges, this paper introduces MixSegNet, a novel deep learning model that enhances crack recognition and segmentation by integrating multi-scale features and deep feature learning. MixSegNet integrates convolutional neural networks (CNNs) and transformer architectures to enhance the detection of small cracks through the extraction and fusion of fine-grained features. Comparative evaluations against mainstream models, including LRASPP, U-Net, Deeplabv3, Swin-UNet, AttuNet, and FCN, demonstrate that MixSegNet achieves superior performance on open-source datasets. Specifically, the model achieved a precision of 95.2%, a recall of 88.2%, an F1 score of 91.5%, and a mean intersection over union (mIoU) of 84.8%, thereby demonstrating its effectiveness and reliability for crack segmentation tasks.

Conventional methods for the early fusion of multi-modal features cannot recognize the relevant modality corresponding to the demand of each user in sequential recommendation. In this paper, we propose the adaptive multi-modal bidirectional long short-term memory network (AM-Bi-LSTM) to recognize the relevant modality for sequential recommendation. Specifically, we construct a new recurrent neural network model that is based on the bidirectional long short-term memory network and obtains multi-modal features, including each user’s sequential actions. Our new modality attention module calculates the importance degree of multi-modal features for sequential operations via the late-fusion approach, which results in the method recognizing the relevant modality. In experiments on a multi-modal and sequential dataset including 14,941 clicks constructed from the largest Web service for teachers in Japan, we demonstrate that AM-Bi-LSTM outperforms existing methods in terms of the diversity, explainability, and accuracy of recommendation. Specifically, we obtain Recall@10 that is 0.1005 better than that of existing early-fusion methods. Moreover, we obtain a value of catalog coverage@10 (representing diversity) that is 0.1710 higher than that for existing methods.

Cross-modal retrieval with noisy labels has attracted much attention. This state-of-the-art method trains a network to increase weights for clean labels in the loss. However, we have found that the network is eventually overfitted to the remaining noisy labels as training progresses. Motivated by this finding, this paper proposes a method called Label Correction using Network prediction based on Memorization Effects (LCNME) to correct noisy labels. This is unlike the state-of-the-art method, which leaves noisy labels on training. We assume that noisy labels are irrelevant to data features and realize label correction using predicted labels (obtained by network prediction) instead of given labels. However, because of memorization effects (the property whereby the network first learns clean labeled data then learns noisy labeled data), predicted labels are contaminated by noisy labels from the certain epoch called the change epoch. Although the change epoch is unknown in advance, we find that it can be identified by observing the loss of the noisy validation set. Using the change epoch, predicted labels can be generated without being affected by noisy labels. Extensive experiments show that LCNME accurately corrects noisy labels and achieves better cross-modal retrieval than existing methods.

In this paper, we propose an image adjustment method for multi-exposure images based on convolutional neural networks (CNNs). We call image regions without information due to saturation and object moving in multi-exposure images lacking areas in this paper. Lacking areas cause the ghosting artifact in fused images from sets of multi-exposure images by conventional fusion methods, which tackle the artifact. To avoid this problem, the proposed method estimates the information of lacking areas via adaptive inpainting. The proposed CNN consists of three networks, warp and refinement, detection, and inpainting networks. The second and third networks detect lacking areas and estimate their pixel values, respectively. In the experiments, it is observed that a simple fusion method with the proposed method outperforms state-of-the-art fusion methods in the peak signal-to-noise ratio. Moreover, the proposed method is applied for various fusion methods as pre-processing, and results show obviously reducing artifacts.

This article presents a method that detects tweet communities with similar topics and ranks the communities by importance measures . By identifying the tweet communities that have high importance measures, it is possible for users to easily find important information about the coronavirus disease (COVID-19). Specifically, we first construct a community network, whose nodes are tweet communities obtained by applying a community detection method to a tweet network. The community network is constructed based on textual similarities between tweet communities and sizes of tweet communities. Second, we apply algorithms for calculating centrality to the community network. Because the obtained centrality is based on tweet community sizes as well, we call it the importance measure in distinction to conventional centrality. The importance measure can simultaneously evaluate the importance of topics in the entire data set and occupancy (or dominance) of tweet communities in the network structure. We conducted experiments by collecting Japanese tweets about COVID-19 from March 1, 2020 to May 15, 2020. The results show that the proposed method is able to extract keywords that have a high correlation with the number of people infected with COVID-19 in Japan. Because users can browse the keywords from a small number of central tweet communities, quick and easy understanding of important information becomes feasible.