GIST Scholar
Not a member yet
30271 research outputs found
Sort by
Activation by Interval-wise Dropout A Simple Way to Prevent Neural Networks from Plasticity Loss
No full textPlasticity loss, a critical challenge in neural network training, limits a model’s ability to adapt to new tasks or shifts in data distribution. This paper introduces AID (Activation by Interval-wise Dropout), a novel method inspired by Dropout, designed to address plasticity loss. Unlike Dropout, AID generates subnetworks by applying Dropout with different probabilities on each preactivation interval. Theoretical analysis reveals that AID regularizes the network, promoting behavior analogous to that of deep linear networks, which do not suffer from plasticity loss. We validate the effectiveness of AID in maintaining plasticity across various benchmarks, including continual learning tasks on standard image classification datasets such as CIFAR10, CIFAR100, and TinyImageNet. Furthermore, we show that AID enhances reinforcement learning performance in the Arcade Learning Environment benchmark. © 2025, ML Research Press. All rights reserved
Advanced post-processing of Ti–6Al–4V alloy fabricated by selective laser melting: A study of laser shock peening and ultrasonic nanocrystal surface modification
No full textIn this study, individual and combined applications of laser shock peening (LSP) and ultrasonic nanocrystal surface modification (UNSM) technologies were applied to Ti–6Al–4V alloy manufactured by the selective laser melting (SLM) process. This study aims to investigate the effects of individual LSP and UNSM technologies, and their sequential combinations in a different order (LSP + UNSM and UNSM + LSP) on the microstructure, grain size, hardness, residual stress, mechanical and tribological properties of Ti–6Al–4V alloy. It is crucial to analyze how the sequential or combined application of LSP and UNSM technologies affects the surface finish of the AM Ti–6Al–4V alloy, which typically has rougher surfaces than traditionally manufactured parts. Results indicated that UNSM technology reduced the surface roughness by about 12 times, introduced a high level of compressive residual stress (CRS), and demonstrated better mechanical and tribological performances than LSP technology. Additionally, the combination of UNSM + LSP technologies, in which LSP technology was subsequently performed on the UNSM-treated sample, was found to be more effective than that of the combination of LSP + UNSM technologies, where the UNSM technology was subsequently performed on the LSP-treated sample. © 2025 The AuthorsTRUEsciescopu
Population Dynamics of the Korean Endemic Monotypic Genus Coreanomecon hylomeconoides Nakai Using an Integral Projection Model
No full textCoreanomecon hylomeconoides Nakai is a perennial herb of a monotypic genus, found only in the southern regions of Korea. Its small population size, restricted geographic range, and unique taxonomic status make it more vulnerable to habitat disturbance, climate change, and other anthropogenic impacts than other comparable species. Therefore, understanding population dynamics of C. hylomeconoides is critical for revealing underlying mechanisms of population responses to varying environmental conditions and informing conservation strategies. Here, we characterize the population dynamics of C. hylomeconoides using an Integral Projection Model (IPM) and Life Table Response Experiment (LTRE), based on demographic data collected from 2022 to 2024 at seven sites covering its entire distribution range. Vital rates (survival, growth, and fecundity) were modeled as size-dependent functions to construct the IPMs. We then calculated population growth rates (λ) and elasticities. Furthermore, population outcomes were decomposed into the contributions of each vital rate to determine which parameters were responsible for spatial and temporal variation. Our results showed an overall decrease in the population size of C. hylomeconoides (λ < 1), suggesting that conservation efforts and management are urgently needed. Growth and survival made the largest contributions to the variation in λ, and survival was negatively correlated with summer precipitation, which may reflect physical disturbance caused by the summer monsoon in Korea. The negative correlation between survival and summer precipitation may be attributed to their tendency to inhabit steep slopes near ravines, where individuals were often observed being dislodged and swept away during intense rainfall in the field. In addition, elasticity analysis highlighted that population viability largely depends on medium- and large-sized individuals, implying that actions to enhance their survival against heavy summer rain are necessary as a key component of conservation strategies.
*Acknowledgements: This study was supported by the Korea National Arboretum (Project No. KNA-25-C-12)
Investigating Temporal Stability and Spatial Information for EEG-Based User Authentication: Shallow ConvNet approach
No full textSecuring sensitive information has become increasingly important in the digital age. EEG-based biometric authentication is considered promising due to its resistance to spoofing and reliance on unique physiological traits. This research explores EEG’s potential for user authentication, focusing on temporal stability and spatial information using the SSVEP paradigm and Shallow ConvNet.
Study 1 examined whether EEG features remain stable over time. Using data from 54 participants collected on two different days, a model trained on Day 1 data achieved classification accuracies of 0.66 (Day 1) and 0.44 (Day 2). A strong positive correlation (0.77) between these accuracies suggests some feasibility for cross-day authentication. Additionally, SSVEP task performance showed moderate positive trends with classification accuracy.
Study 2 investigated EEG’s spatial characteristics, suggesting the parietal and occipital regions hold more information for classification than frontal and temporal regions. Masking experiments indicated focusing on these critical regions might reduce channel requirements and improve accessibility for EEG-based systems.
While the findings suggest potential for SSVEP and Shallow ConvNet in EEG authentication, further research is needed to address limitations.|민감한 정보 보안이 디지털 시대에서 점점 더 중요해지고 있다. EEG 기반 생체 인증은 스푸핑 공격에 강하고 독특한 생리적 특징을 활용한다는 점에서 유망한 방법으로 여겨진다. 본 연구는 SSVEP 패러다임과 Shallow ConvNet을 활용하여 시간적 안정성과 공간적 정보를 중심으로 EEG의 사용자 인증 가능성을 탐구했다.
Study 1에서는 EEG 특징이 시간에 따라 안정적으로 유지되는지를 조사했다. 54명의 데이터를 두 날에 걸쳐 수집했으며, Day 1 데이터를 학습한 모델은 Day 1에서 0.66, Day 2에서 0.44의 분류 정확도를 기록했다. 두 날의 정확도 간 강한 양의 상관관계(0.77)가 나타나 날짜를 초월한 인증 가능성을 시사했다. 또한, SSVEP 실험 수행도와 분류 정확도 간 중간 정도의 상관관계가 확인되었다.
Study 2에서는 EEG의 공간적 특성을 분석했다. 두정엽과 후두엽이 사용자 분류에서 전두엽과 측두엽보다 더 많은 정보를 제공한다는 결과가 도출되었다. 마스킹 실험 결과, 주요 영역에 집중하면 필요한 채널 수를 줄여 EEG 기반 시스템의 접근성을 높일 가능성이 있음을 보여주었다.
이번 연구는 SSVEP와 Shallow ConvNet이 EEG 생체 인증에 있어 유망한 잠재력을 가질 수 있음을 시사하지만, 이를 검증하고 한계를 보완하기 위한 추가 연구가 필요하다.MasterAbstract (English) i
Abstract (Korean) ii
List of Contents ii
List of Tables v
List of Figures vi
I. Introduction 1
1.1 Background 1
1.2 Why was SSVEP chosen 2
1.3 Why was Shallow ConvNet chosen 3
1.4 About Study 1, 2 4
II. Study 1: Does EEG feature remain another day 5
2.1 INTRODUCTION 5
2.2 MATERIALS AND METHOD 6
2.2.1 EEG data 6
2.2.2 Classification model 8
2.3 RESULTS 10
2.3.1 Classification result 10
2.3.2 Subject accuracy distribution by day 11
2.3.3 Correlation 13
2.4 DISCUSSION 15
III. Study 2: Which brain region is important 17
3.1 INTRODUCTION 17
3.2 MATERIALS AND METHODS 18
3.2.1 Experiment 18
3.2.2 Preprocessing & Classification model 20
3.2.3 Brain Region Masking 21
3.3 RESULTS 23
3.3.1 Model training 23
3.3.2 Channel shuffling 24
3.3.3 Confusion matrix 24
3.3.4 Accuracy, precision, recall 26
3.4 DISCUSSION 27
IV. Conclusion 31
References 33
A Abbreviations 38
Acknowledgements 3
Connecting tubules: mechanisms of endoplasmic reticulum membrane fusion
No full textAtlastins (ATLs) are integral dynamin-like GTPases that are critical for the formation and maintenance of the endoplasmic reticulum (ER) network, one of the most complex and essential organelles in eukaryotic cells. The ER, which is composed of interconnected tubules and sheets, serves vital functions, including calcium storage, protein and lipid synthesis, and inter-organelle communication. Homotypic membrane fusion, mediated by ATLs, ensures the tubular structure of the ER by generating and stabilizing three-way junctions. Humans express three ATL paralogs, called ATL1, ATL2, and ATL3, which have distinct expression patterns and regulatory mechanisms. Mutations in these proteins are linked to hereditary sensory neuropathies and hereditary spastic paraplegia, highlighting their critical importance in cellular and neuronal health. Here, we review recent studies providing insights into how ATLs are regulated by their N-and C-terminal extensions, as well as how extrinsic factors potentially regulate the activities of ATLs to establish and maintain the normal ER structure.TRUEsciescopu
Subject-Independent sEMG-Based Prosthetic Control Using MAMBA2 with Domain Adaptation
No full textIntegrating functional wrist articulation in prosthetic robot arms is crucial for enhancing natural movement and reducing compensatory upper limb motions. However, two significant challenges remain in electromyography (sEMG)-based prosthetic control: (1) real-time processing via efficient model design and (2) cross-subject generalization to address the individual variability in muscle signals. This study employs the MAMBA2 architecture to address the first challenge, leveraging Structured State Space Models (SSM) for efficient long-sequence inference. This enables real-time control with minimal computational overhead, making it well-suited for prosthetic robot arm applications. To tackle the second challenge, we implement a Representation Subspace Distance (RSD)-based Unsupervised Domain Adaptation (UDA), which preserves feature scale while aligning inter-subject variations, mitigating domain shift effects, and improving subject-independent wrist movement estimation. The model is trained on the Ninapro DB2 dataset, utilizing multi-channel sEMG signals and corresponding wrist kinematics. Evaluation results demonstrate that the MAMBA architecture outperforms conventional recurrent neural networks, achieving lower Mean Squared Error (MSE) and higher R2 values, with the Attention variant exhibiting the best prediction performance. Furthermore, this study highlights that the proposed UDA approach, combined with RSD-based alignment, significantly enhances cross-subject performance, reducing the need for extensive calibration. By enabling real-time processing through a computationally efficient model structure and effectively addressing cross-subject variability, this study contributes to developing a more reliable and generalizable sEMG-based robotic prosthesis controller, ultimately improving its applicability across diverse individuals
Electron and Hole Trapping Characteristics of a Low-Temperature Atomic Layer-Deposited HfO2 Charge-Trap Layer for Charge-Trap Flash Memory
No full textScaling down the charge-trap memory cell for high storage density causes severe reliability issues such as the decreased trapped charge density, migration of stored charges to adjacent cells, electrostatic interference between neighboring cells, and gate dielectric breakdown. Therefore, it is highly required to explore the advanced charge-trap layer (CTL) having a high trap density with a deep level for improved performance and reliability. In this study, nonvolatile charge-trap memory characteristics are demonstrated using a low-temperature atomic layer deposition (ALD) of hafnium oxide (HfO2) CTL and Al2O3 tunneling and blocking oxides. The use of a high-k dielectric stack enhances the electric field for efficient and reliable device operations in scaled-down devices. In particular, the low-temperature ALD HfO2 CTL deposited at 50 °C has a high charge-trap areal density of 9.65 × 1012 cm-2, exhibiting a large threshold voltage shift of ∼5 V. The proposed device presents a nonvolatile retention of 81.7% for 10 h thanks to the amorphous phase of the low-temperature HfO2 CTL, in contrast to a poor retention of 44.8% in the device with the crystalline high-temperature HfO2 CTL deposited at 200 °C. Furthermore, rapid thermal annealing at 600 °C on the dielectric stack significantly enhances hole trapping in the HfO2 CTL via creation of acceptor-level traps by interdiffusion between HfO2 and Al2O3, securing the large threshold voltage shift of ∼7.8 V. It paves the way for providing the optimized gate dielectric stack of CTF consisting of Al2O3 and defective HfO2 for improved CTF characteristics. © 2025 American Chemical Society.FALSEsciescopu
Designing a Lakehouse for Multi-Sensing Digital Twin
No full textDigital Twin is emerging as a core technology of the fourth industrial revolution, but it comes with complex challenges in efficiently integrating and processing data generated from various sources. To solve this problem, data architectures such as data warehouse, data lake, and data lakehouse are being introduced into digital twin systems. In this thesis, we present the design of a lakehouse for efficiently processing multi-sensing data generated in Dream-AI Space, an AI convergence living lab. Its performance was verified through experimental validation. The proposed lakehouse utilizes Medallion architecture and ELT processes to maximize data quality and processing efficiency, and integrates, manages, and synchronizes UWB RTLS data and GPU metric data in an OpenUSD-based Dream-AI Space environment. The Lakehouse's concurrent processing capabilities and time travel functions were designed to support real-time processing of multi-sensing data in Dream-AI Space. Comparative experiments with the lake architecture demonstrated that the lakehouse is effective for Dream-AI Space.|디지털 트윈은 4차 산업혁명의 핵심 기술로 부상하고 있지만, 다양한 소스에서 생성되는 데이터를 효율적으로 통합하고 처리하는 데 있어 복잡한 도전 과제가 존재한다. 이러한 문제를 해결하기 위해 데이터 웨어하우스, 데이터 레이크, 데이터 레이크하우스와 같은 데이터 아키텍 처가 디지털 트윈 시스템에 도입되고 있다. 본 논문에서는 AI 융합 리빙랩인 Dream-AI Space 에서 생성되는 멀티센싱 데이터를 효율적으로 처리하기 위한 레이크하우스 설계를 제안하고, 그 성능을 실험적으로 검증하였다. 제안된 레이크하우스는 데이터 품질과 처리 효율을 극대화 하기 위해 Medallion 아키텍처와 ELT 프로세스를 활용하며, OpenUSD 기반 Dream-AI Space 환경에서 UWB RTLS 데이터와 GPU 메트릭 데이터를 통합, 관리, 동기화한다. 레이크하우스 의 동시 처리 기능과 time travel 기능은 Dream-AI Space의 멀티센싱 데이터를 실시간으로 처리할 수 있도록 설계되었다. 레이크 아키텍처와의 비교 실험을 통해 레이크하우스가 Dream-AI Space에 효과적임이 입증되었다.MasterAbstract (English) i
Abstract (Korean) ii
List of Contents iii
List of Tables v
List of Figures vi
I. Introduction 1
II. Related Work 5
2.1 Digital Twin 5
2.1.1 Concept and Origin of Digital Twin 5
2.1.2 Definition of Digital Twin 7
2.1.3 Domain-specific Data Types for Digital Twins 8
2.2 Approaches for Multi-Sensing Data Integration for Digital Twin 11
2.2.1 Diverse Approaches for Data Integration 11
2.2.2 Challenges in Data Integration 13
2.3 Data Storage Architectures for Data Integration and Management 15
2.3.1 Data Warehouse 16
2.3.2 Data Lake 18
2.3.3 Data Lakehouse 20
III. Multi-Sensing Dream-AI Digital Twin Space 25
3.1 Infrastructure and Clusters in Dream-AI Space 25
3.2 Digital Twin Services in Dream-AI Space 27
IV. Designing a Lakehouse for Dream-AI Space 30
4.1 OpenUSD based Dream-AI Space 30
4.1.1 Hierarchical Dream-AI Space Based on OpenUSD 30
4.1.2 Mapping Multi-Sensing Data to USD Prim in Dream-AI Space 32
4.2 Lakehouse Design 33
4.2.1 Role of Lakehouse for Dream-AI Space 33
4.2.2 Lakehouse Data Workflow for Dream-AI Space 34
V. Validation of Designed Lakehouse 37
5.1 Experimental Setup for Validating the Lakehouse Design 37
5.2 Validation of Lakehouse for Dream-AI Space 38
5.2.1 Comparison of Time Travel Loading Speed in Lake and Lakehouse 39
5.2.2 Concurrency Validation in Lakehouse for Dream-AI Space 41
VI. Conclusion 43
References 4
Exploring acidity-dependent PCET pathways in imino-bipyridyl cobalt complexes
No full textThe electrochemical proton reactivity of transition metal complexes has received intensive attention in catalyst research. The proton-coupled electron transfer (PCET) process, influenced by the coordination geometry, determines the catalytic reaction mechanisms. Additionally, the pKa value of a proton source, as an external factor, plays a crucial role in regulating the proton transfer step. Understanding the effects of variations in the pKa values of Brønsted acids on the PCET process is therefore essential. This study compares the PCET pathways of two high-spin cobalt (Co) complexes with contrasting exchange coupling interactions under acidic conditions with high and low pKa values. These findings reveal how proton reduction reactions in high-spin Co complexes are affected by the internal factor of the spin state, as well as an external factor related to the proton source. The corresponding reaction mechanisms are also proposed based on these observations.FALSEsciescopu
Growth hormone is required for hippocampal engram cell maturation
No full textMemory is stored in specialized cells known as engram cells, but the molecular mechanisms underlying this process remain unclear. We discovered that the initial translation leading to growth hormone (GH) synthesis in engram cells is required for their maturation in the mouse dentate gyrus. Time-course experiments with a translation inhibitor suggested that the key molecule is immediately translated during learning. We identified GH as the crucial factor, specifically translated within the first few minutes of consolidation in engram cells. Blocking GH activity with a dominant-negative (G118R) mutation blocked engram maturation, whereas facilitating activity-dependent GH uptake alleviated engram maturation deficits caused by translation inhibition. Together, our findings propose GH as a key mediator of hippocampal engram cell maturation.TRUEsciescopu