Case Western Reserve University (CWRU) and Tohoku University (TU) have been exchanging academic and research knowledge and experience focused on data science over the past 10 years. During this decade, our exchange activities have been continuing and extending year by year. Our extensive exchanges are characterized by more than ten joint symposia organized alternately in CWRU and TU, research collaborations, and the acceptance of visiting professors. The data science education at TU has been constantly inspired by CWRU, which has been conducting cross-disciplinary data science education for undergraduate as well as graduate students since the very beginning of "the Data Era." Now, resources and know-how of the data science education program are shared with the other graduate schools at TU. Similar growth has been experienced at CWRU in Applied Data Science at the undergraduate and graduate levels, where recently a new graduate certificate was developed with CWRU's Mandel School of Applied Social Sciences focusing on Data Science for Social Impact. In addition to data science education, unprecedented-scale data produced by a large-scale measurement facility such as a synchrotron is a common target for data science research focus of CWRU and TU. Such a coincidence of research could suggest perspectives for our collaborations in the next decade.

In this study, we developed a learning method for constructing a neural network system capable of memorizing data and recalling it without parameter updates. The system we built using this method is called the Appendable Memory system. The Appendable Memory system enables an artificial intelligence (AI) to acquire new knowledge even after deployment. It consists of two AIs: the Memorizer and the Recaller. This system is a key–value store built using neural networks. The Memorizer receives data and stores it in the Appendable Memory vector, which is dynamically updated when the AI acquires new knowledge. Meanwhile, the Recaller retrieves information from the Appendable Memory vector. What we want to teach AI in this study are the operations of memorizing and recalling information. However, traditional machine learning methods make AI learn features inherent in the learning dataset. We demonstrate that the systems we intend to create cannot be realized by current machine learning methods, that is, by merely repeating the input and output learning sequences with AI. Instead, we propose a method to teach AI to learn operations, by completely removing the features contained in the learning dataset. Specifically, we probabilized all the data involved in learning. This measure prevented AI from learning the features of the data. The learning method proposed in the study differs from traditional machine learning methods and provides fundamental approaches for building an AI system that can store information in a finite memory and recall it at a later date.

This paper applies Cox regression models to a dataset from Tokyo rental property market, collected from March 2019 to March 2021. We extend Cox regression models using deep learning to allow liquidity and price elasticity of rental properties to depend on time and location in a nonlinear manner, analyzing the effects of the COVID-19 pandemic on the rental market. Neural networks are employed to model liquidity and price elasticity, maintaining interpretability. Our findings indicate that the pandemic has highlighted and accelerated the trend of liquidity and price elasticity spreading from central Tokyo to the surrounding areas, reflecting changes in demand patterns likely influenced by factors such as remote work and a preference for less densely populated areas.

This study verified the applicability of transfer learning of illustration tag discrimination models across different sites and constructed a tag recommendation system that contributes to increasing the number of views. Using transfer learning, we relearned a previously trained model into a model specialized for pixiv tags and evaluated its performance. We also developed a model for predicting the number of views of illustrations, and proposed a method for recommending tags that increase the number of views by combining it with a tag discrimination model. This method is capable of making recommendations that take into account both the "image suitability" of the tag and its "effectiveness in increasing views," providing new insights into both automatic tagging and engagement promotion.

As a novel way of presenting information, augmented reality (AR) enables people to interact with the physical world in a direct and intuitive way. While there are some mobile AR products implemented with specific hardware at a high cost, the software approaches of AR implementation on mobile platforms (such as smartphones, tablet PC, etc.) are still far from practical use. GPS-based mobile AR systems usually perform poorly due to the inaccurate positioning in the indoor environment. Previous vision-based pose estimation methods need to continuously track predefined markers within a short distance, which greatly degrade user experience. This paper first conducts a comprehensive study of the state-of-the-art AR and localization systems on mobile platforms. Then, we propose an effective indoor mobile AR framework. In the framework, a fusional localization method and a new pose estimation implementation are developed to increase the overall matching rate and thus improving AR display accuracy. Experiments show that our framework has higher performance than approaches purely based on images or Wi-Fi signals. We achieve low average error distances (0.61–0.81m) and accurate matching rates (77–82%) when the average sampling grid length is set to 0.5m.

Bin-picking is a problem of an object to be automatically picked up from a randomly stacked pile. When considering the complex light reflection scenes, Light Transport Matrix (LTM) estimation based 3D measurement method achieves high accuracy and robustness; however, it is computationally expensive. To achieve the bin-picking such a real-time application for complex light reflection scenes, we propose a new learning-based 3D object recognition and pose estimation method. We leverage a neural network for learning features of point clouds in order to detect and estimate 3D position of the object. We develop a deep learning model which is trained by using the synthetic point cloud data. The key idea of our method is to separate translation estimation and rotation estimation, and introduce the attention mechanism to aggregate the pair-wise feature and the point-wise feature. We train the network using the dataset from a simulation, and test this trained network on the real scene. We also integrate the LTM estimation-based 3D measurement and proposed object detection and pose estimaition with a robot system to achieve the bin-picking task.

Sensing the light transport between light sources and optical sensors is a fundamental issue in computational photography. A well-known linear light-transport model, known as the light transport matrix (LTM), includes all pixel pairs between light sources and optical sensors. LTM is a large matrix and is often estimated using sparse estimation. However, in some cases, this linear observation model may not be appropriate for actual observations, because in practice, observations involve non-linearities. To address a specific aspect of this problem, Saturation ADMM ℓ₁ minimization can be used. It works well even if the observation is performed under saturated conditions. In this study, we perform an experiment that reveals another non-linearity: the under-exposure condition, which is when the captured camera intensities are very weak, they sometimes less than noise. Under this condition, the intensity of the projector pixel should be propagated to a camera pixel corresponding to the LTM; however, it may not be captured. We propose an ℓ₁ minimization method for saturation and under-exposure problems that occur simultaneously. Through numerical simulations and experiments using an actual projector-camera system, we evaluate whether our method can perform accurate estimations under such conditions.

Traditional scientific investigations often have poor or non-existent metadata and data management plans, which poses many challenges for the efficiency, transparency, and reproducibility of research studies. Due to these practices, historical data are rarely leveraged in future studies, but rather are frequently forgotten on a local computer or hardware. We describe a data-driven study protocol design to overcome such challenges in research investigations and to garner the maximum value of available and existing data. A study protocol focuses on the importance and lifespan of data beyond a study where data is re-used to generate new insights. We outline management strategies to ensure that data and metadata, including historical results, are properly linked and recorded using the domain-driven ontologies and FAIR (Findable, Accessible, Interoperable, and Reusable) guiding principles.

Next-generation synchrotron facilities provide unparalleled capabilities for material characterization techniques, such as X-ray diffractometry (XRD), wide-angle X-ray scattering (WAXS), small-angle X-ray scattering (SAXS), and high-energy diffraction microscopy (HEDM). However, the rapid growth of data generated at these facilities has created significant challenges in data storage, metadata management, and analysis. Traditional methods struggle to keep pace with the high-throughput data streams, leading to inefficiencies in data processing, accessibility, and metadata management. This paper presents a perspective on synchrotron data science that addresses the critical issues of data deluge, metadata standardization, and the interoperability of experimental data across different beamlines and facilities. We highlight the role of ontologies in structuring, integrating, and enabling principles for making synchrotron data findable, accessible, interoperable, and reusable (FAIR). Also, we propose a road map for implementing ontology-based frameworks and AI-assisted workflows in the Common Research Analytics and Data Lifecycle Environment (CRADLE), a distributed computing platform to enhance the efficiency and scientific impact of synchrotron data analysis.

Electromagnetic levitation combined with high-speed imaging enables direct in-situ observation of crystal growth dynamics in high-temperature melts, providing crucial insights for materials synthesis. However, high-speed imaging generates massive datasets, often exceeding tens of thousands of frames per experiment, which poses significant challenges for traditional manual characterization methods. We demonstrate the evolution from computer vision methods to machine learning approaches using U-Net through three experimental studies of aluminum nitride (AlN) crystal growth in Ni–Al and Fe–Al systems at 1850–2030 K. Our methodological progression from basic computer-vision image processing to machine learning establishes increasingly robust frameworks for quantifying nucleation events, AlN crystal growth rates, and morphological evolution. When applied to electromagnetic levitation experiments with synchronized dual-camera imaging, these automated techniques revealed quantitative relationships between thermodynamic driving forces and crystal orientation that would be impractical to extract manually. Notably, the deep learning approach in this work achieved 95% IoU in detecting crystal formation area while reducing analysis time to minutes. This deep learning framework is applicable beyond crystal growth studies, offering a template for automated analysis in other in-situ characterization techniques where rapid dynamic processes generate substantial datasets.

For using a drug to treat a disease, healthcare professionals need access to the knowledge regarding adverse drug events (ADEs) on a daily basis to shield their patients from unexpected health hazards. Biomedical literature is a beneficial information resource based on its abundance of citations and high update frequency. Previous studies that extracted ADEs from biomedical literature did not focus on the first reports of ADEs, which healthcare professionals should pay attention to.

In this paper, we evaluated ADEs that were reported for the first time in the literature by authors' arguments in the title and abstract with an author novelty score, which we defined. We then developed a method of simple word matching to predict the author novelty score, which increased the likelihood of identifying the first reports of ADEs based on authors' arguments by seven-fold. The predicted author novelty scores were obtained from our database of ADEs. As the first reports of ADEs that we observed consisted of a variety of drugs and adverse events that are physiologically plausible, our database may be useful for healthcare professionals to make decisions regarding specific drug treatment.

Neutrophils are the major populations of white blood cells and have been reported to facilitate cancer metastasis. Meanwhile, emerging evidence has recently suggested the anti-cancer role of neutrophils. Our previous study revealed that CB-839 and 5-FU-treated colorectal cancer (CRC) tumors recruited neutrophils and induced neutrophil extracellular traps (NETs). Cathepsin G (CTSG), which is released during NET formation, enters CRC cells through the receptor for advanced glycation end products (RAGE) and cleaves 14-3-3ε to promote apoptosis. However, the detailed mechanism underlying CTSG's anti-tumor function remains less studied. In this study, we report that CTSG enters CRC cells through RAGE-mediated endocytosis. Knocking out RAGE or inhibiting endocytosis blocks CTSG from entering CRC cells and attenuates CTSG-induced apoptosis. Furthermore, the clathrin coat assembly complex and SNARE proteins were enriched in an arrayed CRISPR/Cas9 screening targeting human membrane trafficking genes. Knocking out SNARE protein STX1A prevents the spread of CTSG in CRC cells and the induction of cleaved PARP. A pooled genome-wide CRISPR/Cas9 screening further identifies the role of CDK1 in the NET-induced killing of CRC cells. Inhibiting CDK1 protected CRC cells from killing by CTSG. Our study reveals novel mechanisms by which CTSG enters and kills CRC cells.