A5102813831- Advanced Neural Network Applications
- Privacy-Preserving Technologies in Data
- Domain Adaptation and Few-Shot Learning
- Ferroelectric and Negative Capacitance Devices
- Adversarial Robustness in Machine Learning
- Stochastic Gradient Optimization Techniques
- Topic Modeling
- Machine Learning in Healthcare
- Cryptography and Data Security
- Advanced Memory and Neural Computing
- Data Management and Algorithms
- Telomeres, Telomerase, and Senescence
- Skin Protection and Aging
- Hormonal and reproductive studies
- Human Motion and Animation
- Robotics and Automated Systems
- Recommender Systems and Techniques
- Traffic Prediction and Management Techniques
- Smart Parking Systems Research
- Age of Information Optimization
- Graph Theory and Algorithms
- Seismology and Earthquake Studies
- Gastrointestinal Bleeding Diagnosis and Treatment
- Machine Learning and ELM
- Data Stream Mining Techniques
Seoul National University
2018-2022
Sungkyunkwan University
2016
Seoul National University Hospital
2009
To promote secure and private artificial intelligence (SPAI), we review studies on the model security data privacy of DNNs. Model allows system to behave as intended without being affected by malicious external influences that can compromise its integrity efficiency. Security attacks be divided based when they occur: if an attack occurs during training, it is known a poisoning attack, inference (after training) termed evasion attack. Poisoning training process corrupting with examples, while...
Skin ageing is caused by numerous factors that result in structural and functional changes cutaneous components. Research has shown senescent cells are known to accumulate skin ageing, however, the role of not been defined.To elucidate cell we evaluated effect senolytic drugs on dermal fibroblasts.Primary human fibroblasts (HDFs) were induced senescence long-term passaging, UV irradiation, H2 O2 treatment. Cell viability was measured after treatment ABT-263 ABT-737 HDFs. Young aged hairless...
Personalized federated learning is aimed at allowing numerous clients to train personalized models while participating in collaborative training a communication-efficient manner without exchanging private data. However, many algorithms assume that have the same neural network architecture, and those for heterogeneous remain understudied. In this study, we propose novel method called classifier averaging (FedClassAvg). Deep networks supervised tasks consist of feature extractor layers....
Making deep neural networks available as a service introduces privacy problems, for which homomorphic encryption of both model and user data potentially offers the solution at highest level. However, difficulty operating on homomorphically encrypted has hitherto limited range operations depth networks. We introduce an extended CKKS scheme MatHEAAN to provide efficient matrix representations together with improved noise control. Using we developed sequential gated recurrent unit called...
The increasing complexity of deep neural networks (DNNs) has made it challenging to exploit existing large-scale data processing pipelines for handling massive and parameters involved in DNN training. Distributed computing platforms GPGPU-based acceleration provide a mainstream solution this computational challenge. In paper, we propose DeepSpark, distributed parallel learning framework that exploits Apache Spark on commodity clusters. To support operations, DeepSpark automatically...
Memory-augmented neural networks (MANNs) are designed for question-answering tasks. It is difficult to run a MANN effectively on accelerators other (NNs), in particular mobile devices, because MANNs require recurrent data paths and various types of operations related external memory access. We implement an accelerator field-programmable gate array (FPGA) based flow architecture. Inference times also reduced by inference thresholding, which data-based maximum inner-product search specialized...
With the recent success of deep learning, amount data and computation continues to grow daily. Hence a distributed learning system that shares training workload has been researched extensively. Although scale-out environment using commodity servers is widely used, not only there limit due synchronous operation communication traffic but also combining neural network (DNN) with existing clusters often demands additional hardware migration between different cluster frameworks or libraries,...
Memory-augmented neural networks (MANNs) were introduced to handle long-term dependent data efficiently. MANNs have shown promising results in question answering (QA) tasks that require holding contexts for a given question. As demands QA on edge devices increased, the utilization of resource-constrained environments has become important. To achieve fast and energy-efficient inference MANNs, we can exploit application-specific hardware accelerators field-programmable gate arrays (FPGAs)....
With the advance in indoor positioning systems such as RFID, WIFI, and Bluetooth, locations of moving objects constitute a significant factor for many applications. Many researches have verified that most people spend their time an environment. In this paper, we propose two novel index structures, called Split Grid Index (SGI) N-Density (N-DSGI) indexing SGI divides grid into set cells when new object enters. particular, if moves cell, it is recursively divided smaller until each cell...
Membership inference (MI) determines whether a given data point is involved in the training of target machine learning model. Thus, notion MI relies on both feature and The existing methods focus model only. We introduce membership disentanglement network (MFDN) to approach from perspective features. assume that features can be disentangled into class are those enable MI, refer trying learn. MFDN disentangles these by adversarial games between encoders auxiliary critic networks. It also...
딥 러닝(Deep learning)은 기존 인공 신경망 내 계층 수를 증가시킴과 동시에 효과적인 학습 방법론을 제시함으로써 객체/음성 인식 및 자연어 처리 등 고수준 문제 해결에 있어 괄목할만한 성과를 보이고 있다. 그러나 학습에 필요한 시간과 리소스가 크다는 한계를 지니고 있어, 이를 줄이기 위한 연구가 활발히 진행되고 본 연구에서는 아파치 스파크 기반 클러스터 컴퓨팅 프레임워크 상에서 러닝을 분산화하는 두 가지 툴(DeepSpark, SparkNet)의 성능을 정확도와 속도 측면에서 측정하고 분석하였다. CIFAR-10/CIFAR-100 데이터를 사용한 실험에서 SparkNet은 과정의 정확도 변동 폭이 적은 반면 DeepSpark는 초기 정확도는 크지만 점차 줄어들면서 SparkNet 대비 약 15% 높은 정확도를 보였고, 조건에 따라 단일 머신보다도 정확도로 보다 빠르게 수렴하는 양상을 확인할 수 있었다.
Various deep learning applications on smartphones have been rapidly rising, but training neural networks (DNNs) has too large computational burden to be executed a single smartphone. A portable cluster, which connects with wireless network and supports parallel computation using them, can potential approach resolve the issue. However, by our findings, limitations of communication restrict cluster size up 30 smartphones. Such small-scale clusters insufficient power train DNNs from scratch. In...
Continuous monitoring of spatial queries has received significant research attention in the past few years. In this paper, we propose an efficient method for continuous top-k queries. contrast to conventional query, a new query considers both and non-spatial attributes. We use grid-based data structure novel grid access method. The proposed (i) quickly identifies moving objects that enter or exit cells (ii) reduces number unnecessary operations by comparing bit-vector information objects....
Interactive drama is a story which requires user`s free choice and participation. In this study, we grasp preference by making training data that utilize characters of interactive drama. Furthermore, describe process implementing systems recommend new users path stories correspond with their preference. We used PCA NMF to extract characteristic The success rate recommending was 75% PCA, while 62.5% NMF.
Distributed training of deep neural networks has received significant research interest, and its major approaches include implementations on multiple GPUs clusters. Parallelization can dramatically improve the efficiency complicated models with large-scale data. A fundamental barrier against speedup DNN training, however, is trade-off between computation communication time. In other words, increasing number worker nodes decreases time consumed in while simultaneously overhead under...
Memory-augmented neural networks (MANNs) are designed for question-answering tasks. It is difficult to run a MANN effectively on accelerators other (NNs), in particular mobile devices, because MANNs require recurrent data paths and various types of operations related external memory access. We implement an accelerator field-programmable gate array (FPGA) based flow architecture. Inference times also reduced by inference thresholding, which data-based maximum inner-product search specialized...