|
|
On Variational Bayes Approach to State Space
Modeling with Its Implementation for Simple
Mathematical Models
Norikazu Ikoma, Nippon Institute of Technology, Japan
Abstract:
Deep Markov Model, which would be called as “Deep Kalman filters”, as
well as “Structured Inference Networks”, models structure behind time
series data by employing nonlinear mapping of neural network for system
model and observation model within state space modeling framework.
So obtained hidden state estimate becomes distributed representation
within neural network that leads to difficulty for interpretation of its
meanings. This work begins with applying simple mathematical models
to the framework of Deep Markov Model in order to address the above
issues. Its implementation employs PyTorch based framework “Pyro”
within programming language Python in demonstrative examples of nu-
merical experiment. The most simple state space model is so-called trend
model has been implement within the framework and parameters have
been estimated via variational Bayes in the numerical experiment.
Contribution of the Paper: Bridge Deep Markov Model and mathe-
matical state space models.
Full Paper (in PDF)
|
IntegraDenoNet: A Deep Learning Based Single Cell
Multiomics Integration and Cell Type Identification
Md Shaharia Hossen, Sakib Mahmood Saad, Maria Akter Rimi, Marin Akter, Fahim Hafiz, Riasat Azim
Abstract:
The correlation between different phases of biological data, such as transcriptomics, metabolomics, and other omics, is important in the case of disease analysis. Multiomics aims to combine diverse omics data into a unified dataset, revealing interrelationships and their influence on complex biological processes. Although multi-omics methodologies are relatively new, their demonstrated potential to accurately uncover insights has captured the bioinformatics field. However, limited datasets and challenges in preparing unbiased models have hindered widespread application. This research introduces an innovative deep learning-based method for the seamless integration of multi-omics single-cell data, allowing for accurate classification of omics expression levels. Omics data are reconstructed using a denoising autoencoder with a learning rate scheduler, cosine annealing. Reconstructed data are integrated with labels for further downstream analysis. Our proposed method achieved minimal classification loss, approximately 0.05% compared to other recent methods. Furthermore, the proposed method achieved a consistent accuracy greater than 90% in three multi-omics datasets, beating four advanced state-of-the-art (SOTA) methods. The proposed model 'IntegraDenoNet' demonstrates improved classification accuracy and advances possibilities in precision medicine.
Contribution of the Paper: 'IntegraDenoNet' leverages deep learning to integrate multi-omics data for cell type classification, achieving 90% accuracy across three gold-standard datasets and outperforming four state-of-the-art methods.
Full Paper (in PDF)
|
Deep Learning-based Post-earthquake Structural
Damage Classification with Small Datasets
Kaveesh Guwanindu ABEYSURIYA, Mihaela Anca CIUPALA, Meghana MANIKONDA, Kruttika JAMALPURAM,
Aishwarya Nitin SONAR, Mhd Saeed SHARIF, Seyed Ali GHORASHI,
Alper ILKI
Abstract:
This paper introduces a deep learning-based approach for automated
classification of local structural element failure modes in post-earthquake
buildings using image-based data. Addressing the critical challenge
of limited training datasets in the structural/earthquake engineering
domain, targeted, domain-informed data augmentation and synthetic
data generation techniques are proposed to enhance dataset size and
diversity. The model architecture and preprocessing pipeline are explicitly
designed to capture damage-sensitive features in images that are essential
for informed decision-making on structural integrity of the building, thus
extending beyond conventional classification tasks. Dataset enhancement,
transfer learning and model regularisation techniques are integrated to
ensure alignment of model predictions with expert domain judgement.
Achieving 0.93 (93%) accuracy, precision, recall and F1-score, the
developed model exhibits robust generalisability without overfitting,
demonstrating clear potential for practical deployment in disaster
resilience and infrastructure recovery efforts.
Contribution of the Paper: The development of a tailored deep
learning-based approach that integrates domain-informed dataset
enhancement, transfer learning and targeted regularisation techniques,
specifically designed to reliably classify local structural element failure
modes from limited, image-based post-earthquake data to support
expert-informed structural integrity decisions.
Full Paper (in PDF)
|
Transfer Learning in Classifying Acoustic Emission Signals from Osteoarthritic Knees
Nazmush Sakib, Md Abdur Rahman, Tawhidul Islam KHAN,
Md. Mehedi Hassan, Shuya Ide
Abstract:
Acoustic emission (AE) is a non-destructive evaluation (NDE) method that allows to inspect the internal condition of material by analyzing the signal which are produced due to the internal change in the condition. Compared to the present methods, due to simplicity and immense potential, AE has gained attention in knee health assessment. With the advancement of computational power many researchers have implemented advanced machine learning (ML) algorithms to characterize the AE signals which were generated from human knees. However, most of this research are focused on implementing the unsupervised ML algorithms. The minimal variability between the AE signals from different knee conditions has posed significant challenges in implementing supervised ML algorithms which shows the promise to make the diagnosis significantly simpler than the present approaches. Therefore, this work aims at implementing transfer learning using CNN and wavelet-based images to classify the AE signals which were generated from the knees of the knee osteoarthritis of different Kellgren Lawrence (KL) grades. VGG-16 CNN model has been trained on the images which were generated from AE signals of the participants. The results shows huge promise of transfer learning in classifying the AE signals from different knee health conditions
Contribution of the Paper: This paper shows a novel application of CNN based transfer learning in diagnosis of knee OA from acoustic emission signals.
Full Paper (in PDF)
|
Acoustic Emission Tomography for Damage Visualization in Homogeneous Material Surface Using Iterative Reconstruction Algorithm
Md Abdur Rahman, Nazmush Sakib, Md Mehedi Hassan, Tashiro Hibiki,
Tawhidul Islam Khan
Abstract:
Acoustic Emission Tomography (AET) is an advanced non-destructive evaluation (NDE) technique that enables subsurface defect detection by analyzing acoustic wave propagation characteristics. This study presents an application of AET for damage visualization in an aluminum specimen using an Iterative Reconstruction Algorithm (IRA). A single acoustic emission (AE) transducer was utilized with 16 linear projections at different angles, employing a dedicated R15α receiver and a transmitter transducer for data acquisition. The investigation was conducted on both an undamaged aluminum surface and an aluminum specimen with an induced circular defect to compare the slowness distribution. The artificially generated signals collected from multiple projection angles were processed to determine the time-of arrival (TOA) variations, which served as the primary input for reconstructing slowness distribution maps. The tomographic reconstruction results demonstrated a uniform slowness distribution in the undamaged aluminum sample, whereas the damaged region exhibited localized increases in slowness, effectively highlighting the internal defect. The ability of AET to visualize subsurface damage through variations in wave propagation characteristics underscores its effectiveness as a diagnostic tool for material integrity assessment. Findings in the research successfully achieved the presence of damage in the material surface which have been shown by changing the slowness mapping. The results highlight the potential of AET for applications in aerospace, automotive, and industrial manufacturing sectors, where real-time structural health monitoring (SHM) and defect detection are critical for ensuring safety and reliability.
Contribution of the Paper: Results indicate the advancement of Tomography for damage detection and visualization in materials using Acoustic signals with integrating an IRA to enhance defect visualization accuracy.
Full Paper (in PDF)
|
MNMD: A multimodal non-invasive mental disorder detection method
Rashed Mustafa, Mahir Shadid, A. H. M. Sajedul Hoque
Abstract:
Millions of people worldwide suffer from mental health conditions like anxiety, stress, and depression, but early and precise detection is still difficult. This study introduces MNMD (Multimodal Non-Invasive Mental Disorder Detection), a system that uses an effective late fusion technique to combine results from the DASS-21 questionnaire with facial expressions in a unique way. In contrast to earlier methods, MNMD focuses on a low-complexity, non-invasive design that combines projected outputs using a "common-string" technique, lowering computational overhead while improving data resilience and variation. The system uses a variety of machine learning models and deep learning frameworks in addition to extensive image feature extraction using Gabor filters and facial landmark detection. With an impressive 98.43% accuracy rate, MNMD offers a quicker, privacy-preserving method of early mental health diagnosis while also demonstrating enhanced prediction performance and practicality for real-world implementation.
Contribution of the Paper: The primary contribution is the development of a novel, non-invasive multimodal fusion method called MNMD, which combines textual and visual data using a low-complexity late fusion technique and achieves higher accuracy (98.43%) in the detection of stress, anxiety, and depression.
Full Paper (in PDF)
|
Hyperventilation-induced Cerebral Patterns
Analysis for Neurocognitive Disorder Detection via
AI Models
Kusum Tara, Ruimin Wang, Yoshitaka Matsuda, Satoru Goto, Takako Mitsudo, Takao Yamasaki, Takenao Sugi
Abstract:
Monitoring hyperventilation (HV)-induced cerebral patterns offers a promising approach for early detection of mild cognitive impairment (MCI) and Alzheimer’s disease (AD), which are associated with disrupted cerebral blood flow, neuronal excitability, and brain connectivity. This study analyzed EEG signals collected under HV-induced physiological stress using deep learning convolutional neural network (CNN)-based AI models to observe brain activity changes. Phase-amplitude coupling (PAC) and spectral topographic mapping (STM) images were generated from EEG and applied to two CNN architectures, MobileNetV2 and Xception. Functional connectivity was further examined using Pearson correlation and Granger causality to identify neural alterations. Among these AI models, Xception combined with PAC images achieved the highest classification accuracy of 98.95%, outperforming MobileNetV2 (96.45%) by effectively capturing non-linear brain dynamics, impaired phase synchronization, and disrupted neural communication in MCI and AD patients. Gradient-weighted class activation mapping (Grad-CAM) was used to rank EEG channels based on their importance, revealing that parietal and occipital channels contributed most to model decisions. These results demonstrate that the proposed PAC-based Xception model provides a reliable method for identifying neurocognitive dysfunction and neural desynchronization of MCI and AD patients with reduced delta-alpha and theta-alpha couplings.
Contribution of the Paper: This study incorporates Granger causality for analyzing HV-induced functional connectivity changes between EEG channels and phase-amplitude coupling with AI models to detect impaired phase synchronization and disrupted neural communication, enhancing reliability in classifying MCI and AD.
Full Paper (in PDF)
|
An Ensemble Approach to Named Entity Recognition for Bangla-English Code-Switched Texts Using XLM-R, BiLSTM-CRF, and CRF
Sultana Tasnim Jahan, Rashed Mustafa
Abstract:
Named Entity Recognition (NER) in Bangla-English code-switched text is quite challenging due to tokenization irregularities, linguistic variances, and the absence of annotated datasets. Current NER algorithms focus mainly on monolingual or organized multilingual texts, with little attention paid to code-switched data. The efficacy of conventional methods is further constrained by the difficulties in managing grammatical errors, unclear entity boundaries, and tokenization issues.
This paper suggests an ensemble-based NER method that combines the CRF, BiLSTM-CRF, and XLM-R models to solve this problem. Entity labels are predicted independently by each model, and a fourth component—an ensemble model that combines the three - is also present. A majority vote process among these four sources determines the final projections. By combining sequence-labeling techniques with transformer-based contextual embeddings, this hybrid approach enhances generalization and lowers recognition mistakes.
Our work shows empirical improvements through extensive experimentation and architectural integration and, in contrast to usual surveys, presents a comprehensive, functional pipeline.
The results of experiments show that the suggested method greatly enhances entity detection in intricate, code-switched texts by achieving more accuracy and robustness when compared to individual models. The study has significant applications in social media analysis, customer support automation, and multilingual information extraction, all of which depend on the ability to handle mixed-language text.
This study confirms performance using both comparison benchmarks and real-case sentence structures from the newly released dataset, in contrast to previous work that assessed code-switched NER separately. We intend to investigate self-learning strategies for domain adaptability, add more varied linguistic patterns, and enlarge the dataset in subsequent research. Furthermore, the performance of NER in code-mixed, low-resource environments may be further improved by using meta-learning techniques and adaptive fine-tuning.
Contribution of the Paper: A newly created Bangla-English code-switched NER dataset is presented in this research along with a novel ensemble-based approach that combines CRF, BiLSTM-CRF, XLM-R, and an ensemble of these models under a single majority voting framework. Accuracy and stability are improved across linguistically inconsistent code-switched texts by combining syntactic, sequential, and contextual modeling.
Full Paper (in PDF)
|
Refining NeRF: The Power of High-Resolution Omnidirectional Vision
Sho Hasegawa, Chinthaka Premachandra
Abstract:
As NeRF technology becomes more widely adopted, research has increasingly focused on enhancing its performance. Despite its potential, NeRF still faces several challenges, including long training times, high computational demands, and lower accuracy compared to other 3D reconstruction methods such as photogrammetry. A critical step in NeRF generation is camera pose estimation, which typically involves extracting features such as object boundaries and corners from captured images. We found that using an omnidirectional camera can reduce shooting time while still enabling accurate NeRF generation, even when the camera lacks a high-performance image sensor. In this study, we aimed to improve the quality of camera pose estimation in order to enhance the accuracy of NeRF generation by increasing the resolution of partitioned omnidirectional images and improving the definition of object boundaries. Our experiments demonstrated that these improvements effectively reduced noise in the generated NeRFs and improved their overall accuracy. Therefore, our findings suggest that even with consumer-grade devices, such as general omnidirectional cameras, it is possible to generate a more accurate NeRF space by incorporating the proposed processing.
Full Paper (in PDF)
|
A Tracking Method of Multiple Animals Using YOLOv5
Toshifumi Kimura,
Hidetoshi Ikeno,
Mizue Ohashi,
Ryuichi Okada,
Mamiko Ozaki,
Hiroyuki Ai,
Shunya Habe,
Teijiro Isokawa
Abstract:
In behavioral experiments on social animals in ethology, it is important to understand not only the detailed location of each of the target animals, but also their swarm behavior emerging from their interactions.
Recently, many systems have been developed to support animal behavior analysis.
However, these systems require videos under good conditions for easy discrimination of targets from background, leading to their limited applications.
In this paper, a tracking system that is robust to different experimental conditions is proposed.
The proposed system adopts YOLOv5, a deep neural network based system, as an object detector from video images and incorporates the existing K-Track system for tracking the detected objects.
The performance of the proposed system is evaluated using actual videos obtained from behavioral experiments, and robust detection of target animals and their tracking is possible.
Contribution of the Paper: Accurate animal tracking is achieved through the combination of object detection and bi-directional associations for detected objects.
Full Paper (in PDF)
|
Realizing Human-Robot Cooperative
Rope-Spinning with Central Pattern
Generator-Based Control Using Visual Information
Kakeru Yamasaki,
Koki Iida,
Patrick Henaff,
Tomohiro Shibata
Abstract:
Achieving coordinated motion through flexible objects remains a significant challenge in Human-Robot Interaction (HRI). This study investigates a novel application of Central Pattern Generator (CPG) control, previously used in handshake robots, to a rope-spinning task involving human-robot cooperation. A real-time motion feedback system was developed using Azure Kinect, enabling a robot to synchronize its movements with human input by dynamically adjusting CPG outputs.
We evaluated the system’s performance by varying rope lengths (250--400 cm) and analyzing spatial trajectories and Euclidean distances between the human and robot end-effectors. Results showed that while high coordination was achieved under shorter rope conditions, longer ropes introduced increased slack and tension variability, which reduced the robot's tracking stability. Frequency analysis also revealed weaker synchronization on the robot side, particularly in the vertical (Z) direction.
These findings indicate that vision-based feedback alone is insufficient for robust adaptation to the dynamic characteristics of flexible objects. The vision-based method demonstrated lower amplitude fidelity and synchronization precision than our previous force-feedback approach. Future work will focus on integrating multimodal feedback, combining visual and force sensing, to improve coordination and robustness in flexible-object-mediated HRI.
Full Paper (in PDF)
|
Data-Driven Diagnosis: Feature Engineering and Hyperparameter Tuning for Imbalanced Cardiovascular Disease Classification
Ankon Karmokar,
Md. Manowarul Islam, Arnisha Akhter, Uzzal Kumar Acharjee
Abstract:
Cardiovascular diseases (CVDs) have become the leading cause of death all over the world, emphasizing the need for prediction models to ensure accurate and timely medical interventions. In this paper, we propose a powerful end-to-end machine learning system that incorporates feature engineering, hyperparameter tuning, and ensemble model analysis to enhance the predictive performance of CVD. Preprocessing encompasses feature engineering methods designed to improve data quality, remove outliers, cap values, and normalize data. Five additional complex models—Random Forest (87.43%), Gradient Boosting (88.07%), AdaBoost (87.01%), XGBoost (88.11%), and LightGBM (88.02%) - are fine-tuned using the RandomizedSearchCV innervation library. XGBoost achieves the highest validation accuracy (88.11%) and is the most effective classifier. The proposed method offers a significant advantage over the conventional one in terms of precision, sensitivity, and F1-score, which can be applied to the screening, prevention, and clinical decision-making in cardiovascular healthcare.
Contribution of the Paper: Key contributions are that data quality can be improved through sophisticated feature engineering, model performance through hyperparameter optimization, and model interpretability via SHAP analysis for better decision-making, especially in sensitive areas such as healthcare.
Full Paper (in PDF)
|
AI-Enabled IoT System for Continuous Health Monitoring and Early Risk Detection
Md. Kamran Sharafi, Mohsinur Rahman Murad, Md. Hasan Mahmud Tonmoy Shak, Md. Manowarul Islam,
Arnisha Akhter, Uzzal Kumar Acharjee
Abstract:
Traditional healthcare systems depend on manual tracking and periodic check-ups, which can postpone the identification of serious health issues particularly for patients with chronic conditions they frequently have difficulty providing timely intervention. In response to this issue, we created a smart health monitoring system that leverages AI and IoT technology to provide ongoing, real-time health assessments and prompt reactions. The system incorporates the ESP32 microcontroller for data handling, as well as sensors like the MAX30100 for measuring blood oxygen and heart rate, the LM35 for body temperature, and an ECG sensor for cardiac monitoring. These sensors gather real-time health data that is sent to a hospital Web server for analysis, risk assessment, and alert notification. With a mobile app, patients can access their health information, get medication reminders, and connect with healthcare providers. By examining the gathered data for unusual patterns and forecasting health risks, AI predictive algorithms facilitate preemptive action. A health monitoring system uses Random Forest and SVM models to consider various physiological data for patient status classification with an accuracy of 99.7–100.0%, aiding in early risk prediction and proactive intervention. Dashboards for healthcare professionals' real-time decision-making, driven by Power BI. This collection of solutions provides patients and caregivers with timely alerts, predictive analytics, enhanced patient involvement, and the capability to monitor wellness around the clock. The system enhances the efficiency of healthcare and the quality of patient care through rapid data-driven responses.
Full Paper (in PDF)
|
Optimizing Task Offloading in Fog Computing with HAGSA-NS: A Hybrid Adaptive Gravitational Search Algorithm
Md. Emran Biswas,
Tangina Sultana, Md. Delowar Hossain, Mst. Khadeja Sarker, Ga-Won Lee, Eui-Nam Huh
Abstract:
The swift proliferation of Internet of Things (IoT) devices and the demand for minimal latency applications has rapidly increased the incorporation of fog computing alongside traditional cloud computing. Nevertheless, efficient task offloading in fog settings remains a significant issue due to variable network conditions, resource constraints, and stringent quality-of-service (QoS) requirements. This research proposes a novel Hybrid Adaptive Gravitational Search Algorithm with Neighborhood Search (HAGSA-NS) to enhance task offloading in fog computing, solving these issues. HAGSA-NS integrates the global exploration capabilities of the Adaptive Gravitational Search Algorithm (AGSA), the diversity-enhancing features of Differential Evolution (DE), and the local exploitation benefits of Neighborhood Search (NS). This hybrid approach enables effective and efficient task allocation, even in highly dynamic and resource-constrained environments. The effectiveness of HAGSA-NS is evaluated against two prevalent optimization methods, Particle Swarm Optimization (PSO) and Genetic Algorithm (GA), using two primary metrics: ideal fitness values across iterations and average computational latency. Experimental results demonstrate that HAGSA-NS consistently outperforms PSO and GA, achieving reduced fitness values and significantly decreased delays. HAGSA-NS achieves an average delay of 1.0, whereas PSO exhibits a delay of 4.234 and GA demonstrates a delay of 1.791. These findings highlight the benefits of HAGSA-NS in terms of solution quality and computational efficiency, positioning it as a viable strategy for work offloading in fog computing environments.
Full Paper (in PDF)
|
Phishing Detection Using Gradient-Weighted Ensemble with Hybrid Sampling and Weighted Ensemble Feature Selection
Mrinal Basak Shuvo, Arjon Talukder, Sadia Islam Neela, Prakriti Paramarthi Roy, Tangina Sultana, Md.
Delowar Hossain, Md. Arshad Ali, Eui-Nam Huh
Abstract:
Phishing attacks lead to a significant cybersecurity threat, where users get tricked with URLs and attackers extract sensitive information. As a result, the attackers gain an advantage and subsequently inflict significant damage. Hence, we have proposed a Machine Learning pipeline that facilitates the reduction of impact caused by phishing attacks. This study explores a competent architecture using various ML classifiers, including Random Forest, XGBoost, and LightGBM. We have used hybrid sampling involving SMOTE, TomekLinks and ADASYN, termed as Triad Sampling Fusion (TSF). For feature selection, we have used a weighted ensemble technique combined with Boruta, RFE, and Elastic Net, referred to as Boruta-RFE-ElasticNet Feature Selector (BREN-FS). For the enhancement of the performance of our model, we have performed GridSearchCV for Hyperparameter Tuning and developed a novel custom-weighted ensemble approach, termed as the Gradient Unified Weighted Ensemble (GUWE), that combines predictions from multiple models using gradient descent to optimize weights dynamically in order to increase classification performance. For training and evaluation, we have used the Mendeley Phishing Dataset. Our used dataset consists of 88,647 instances along with 111 features. Existing studies retain a relatively large feature subset, which increases computational complexity and overfitting risks. By applying TSF, BREN-FS, and comprehensive evaluation approaches such as GUWE, our study achieved an accuracy of 97. 52\% that improves the effectiveness of phishing detection, significantly contributing to cybersecurity and reducing online threats.
Full Paper (in PDF)
|
A Real-Time Vision-Based System for Human Gesture Recognition in Collaborative Work Cells
Natchanon Suppaadirek,
Shreyan Shukla,
Piyush Mudgal,
Tomohiro Shibata
Abstract:
Human-robot interaction is a major field of investigation, focusing on the optimization of working processes as well as employee productivity. Despite an enormous amount of progress made in this direction, the necessity to develop systems targeting people with disabilities remains a pressing need. This report presents a novel paradigm for assistive robotics via the development of an intelligent work cell for aging individuals and physically disabled persons. The system merges depth camera technology, light machine learning, and MediaPipe-based human tracking to enable real-time human-robot interaction through accurate inference of user intent. Key innovations include a gimbal-mounted depth camera for motion tracking of the user, a modular 3D-printed gripper for easily customizing manipulation, and an efficient gesture classification pipeline. Experimental results demonstrate that the system achieves over 90% average gesture recognition accuracy, which is comparable to or higher than similar gesture-based systems, with real-time performance. The system bridges the gap between theoretical research and practical application in assistive robotics.
Contribution of the Paper: The contribution of this paper lies in the development and evaluation of an innovative system to interpret human gestures through vision-based technology, with a specific focus on improving human-robot collaboration.
Full Paper (in PDF)
|
Prediction of QoL in healthy older adults using non-motor information from smart devices
Yu Makido, John Noel Victorino, Kengo Iwamoto, Tomohiro Shibata
Abstract:
Although global interest in well-being and QoL is increasing, continuous awareness of one's QoL in daily life remains challenging due to the need for repeated questionnaire responses. In this study, we evaluate the performance of a prediction model for QoL in healthy older adults using a Garmin Venu 3S fitness tracker and a FonLog data collection application to collect non-motor information and QoL data, and predict QoL using a support vector machine (SVM). The results of the prediction using a SVM showed that the Accuracy was approximately 0.96 and the F1-Score for each class was approximately 0.88 or higher. These results suggest the effectiveness of the QoL prediction model using non-motor information. In the future, we plan to improve the processing and prediction in real time, and to evaluate the accessibility, usability, and effectiveness of the system for a wider range of users through experiments with non-motor subjects.
Contribution of the Paper: Since smartwatches can be worn easily and are not difficult to use on a daily basis, the development of a prediction model is expected to be an easy way to measure QoL.
Full Paper (in PDF)
|
Gaze analysis of a walker user for the development of a gaze-based interface to operate a robotic walker
Kengo Iwamoto, Yuuta Shinoda, Tomohiro Shibata
Abstract:
As populations in advanced economies continue to age and birth rates decline, a growing shortage of caregivers has emerged.
This shortage has led to an inability to meet the demand for rehabilitation through human caregivers, prompting research into the automation of rehabilitation, such as robotic walkers. Estimating the user's intent in a robotic walker can improve safety and provide intuitive control, as well as personalized assistance, thus reducing the psychological barriers users may face when interacting with the robot. This study aims to investigate the intention of direction change based on head orientation by analyzing gaze patterns during turning and while checking the surroundings. Gaze analysis was performed using the Tobii Pro Glasses 3. Participants were asked to perform a task involving both turning while navigating a specific route and checking numbers placed around them, allowing for the collection of gaze data. The collected gaze data were analyzed using three machine learning models: Random Forest, LightGBM, and SVM—which are capable of handling high-dimensional datasets and are expected to achieve high classification performance. Using gaze data collected during surrounding check tasks and direction change tasks, a classification model was trained to distinguish between surrounding check behavior (Class 0) and direction change behavior (Class 1). As a result, the Random Forest model achieved a classification accuracy of 99.5%, the LightGBM model 99.8%, and the SVM model 99.4% for healthy participants, consistently demonstrating high accuracy. For patients with Parkinson's disease (PD), the model trained on healthy participants could not be directly applied. Still, an attempt to improve SVM classification accuracy by adjusting the threshold using the decision function resulted in a classification accuracy of 64% at a threshold of 0.5.
Contribution of the Paper: This study's main contribution is verifying the effectiveness of gaze-based classification for distinguishing between turning and checking the surroundings in healthy individuals and PD patients, concerning head orientation.
Full Paper (in PDF)
|
Quantum computation as trajectory monitoring
requires only one qubit in answer register in
quantum phase estimation
Teturo Itami, Nobuyuki Matsui, Teijiro Isoka
Abstract:
Quantum computation scheme developed using classical apparatuses has
been extended to systems with two qubits. To obtain controlled NOT
operation, we set an interaction between two bits, control bit and target bit, with an oscillation on the target bit. Condition of controlled
NOT operation simultaneously determines the operation time, interaction strength and oscillation magnitude. Subsequently, we demonstrate
that, in quantum phase estimation algorithm, our system — with only
one qubit in answer register — requires no iteration. Classical mechanical systems we apply do not share the same vulnerabilities as quantum
systems. Moreover, our system does not require numerous measurements
to determine most probable value of the system output. These features
provide compelling advantages for such computational systems.
Contribution of the Paper: Our work shows a concrete method to obtain controlled NOT operation between two qubits. The result provides a
method for assembling classical apparatuses to construct universal gates
in quantum computers.
Full Paper (in PDF)
|
|
