Research Projects


연구과제 목록

1. [SK Hynix], SK Hynix  向 생성형 AI 구축에 대한 자문, 2024/04/03 ~ 2024/10/02

2. [삼성디스플레이], Deep Learning 기반 불량 혐의설비 탐지 모델 개발, 2024/04/01 ~ 2025/03/31

3. [온택트헬스], 의료 AI 소프트웨어 기술개발 자문, 2024/03/01 ~ 2025/02/28 

4. [코난테크놀로지], Advanced Domain Adaptation 솔루션 개발, 2023/12/18 ~ 2024/06/17

5. [한국연구재단 - 우수신진], 제조 공정 진단 일반화 성능 개선을 위한 지식융합 인공지능 방법론 개발, 2023/03/01 ~ 2026/02/28

6. [연세대학교 - 미래선도], 이질성이 있는 데이터 분석을 위한 지식융합 인공지능 학습 방법론 개발, 2021/11/01 ~ 2024/10/31

7. [티라유텍], 패션 드레이핑 자동화를 위한 의복 패턴 매칭 알고리즘 개발, 2024/01/29 ~ 2024/06/30

8. [SK Hynix], CD-SEM 장비 관리 고도화를 위한 Image 기반 HW Reponse Para 발굴, 2023/04/24 ~ 2024/04/23 (종료)

9. [LG에너지솔루션], 도메인적응을 통한 테스트베드 모델의 실제 설비 적용, 2023/04/01 ~ 2023/08/31 (종료)

Task-Agnostic Unified Anomaly Detection

1. Develop a robust unified anomaly detection framework capable of handling various anomaly detection tasks with a single architecture  

2. Explore self-supervised representations  for unified anomaly detection

3. The framework incorporates techniques such as BPM and top-k ratio feature matching, enabling task unification in a task-agnostic manner 

Multi-modal Few-Shot Learning

1. Addressing challenges in manufacturing, such as limited labels, drift, lack of references, and device changes 

2. Utilizing various types of inputs (Image, Sensor data, Structured data, Text, etc) to identify better representations

3. Advancing towards the development of general-purpose Artificial Intelligence

Robust Deepfake Disruption in Real-World Scenario

1. Our focus is on attacking the latent encoding process without relying on specific target attributes

2. We are designing an ensemble strategy with high scalability to target deepfake models, encompasses all three categories of deepfakes, including both GAN-based and Diffusion-based models

3. Our approach disrupts real-world scenarios, ranging from white-box, gray-box, and even black-box scenarios, where specific deepfake models are unidentified

Knowledge-informed Machine Learning

1. Integrating domain knowledge into the training process, leads to improving regular machine learning models for addressing limited samples

2. Encoding different types of modalities into the framework

3. Uncertainty Quantification

Anomaly Detection 

1. Limited supervision but powerful vision inspection framework 

2. Building a new type of anomaly detection benchmark dataset ‘AutoVID’ 

3. Proven excellence through various experiments with MVTecAD, Oxford Flowers 102, and AutoVID  (applicable to detect subtle process defects as well), compare to PatchCore

 SimVI* (A Simple Visual Inspection Framework with foreground object segmentation and pre-trained model-based anomaly detection)  

5.  Openset Anomaly Detection

Unsupervised Domain Adaptation 

1. Cluster-weighted and class-informed adversarial learning, Negative / Positive sample identification, Source structural regularization

2. Generate domain invariant features for improving performance and providing reliable results.

Other Ongoing Research

Explainable Deep Learning on Multimodal Data for Boiling Crisis in Nuclear Reactors

1. Leverage multimodal data to ensure the model robustness and reliability. We are developing a heterogeneous data fusion technique to utilize the joint power of data from multi-modality

2. Extract hidden physics from modality-dependent features

Building Fault Detection Baseline Construction

1. Construct baselines by sampling from the multiple sensor readings collected from real building systems for fault test cases

2. A model-free decision metric, relying on the data to extract information on sample sufficiency and characterize heterogeneity of a dataset

Imaging-based Diagnosis

1. Feature Transfer Enabled Multi-Task Deep Learning Model offers 1) a fully automatic system handling detection, segmentation, and classification, 2) cross-view features transferring for improved model performance, and 3) a lower risk of negative transferring issues (transfer only within the same domain) 

2. Deep Residual Inception Encoder-Decoder Network for Medical Imaging Synthesis generates synthetic images by utilizing knowledge transfer between image modalities 

Sensor-based patient diagnosis for fall risk using a 10-meter walking test

Predicting fall risks among community-dwelling older adults using linear and nonlinear gait variability features and random decision forest framework

Smartphone-based Telemonitoring of Parkinson's Disease Patients

1. Developing the telemonitoring framework to leverage other patient's beneficial information (avoiding negative transfer) when building a predictive model for a target patient

2. Predicting severity of Parkinson's Disease from smartphone-based features using feature + instance selection in a semi-supervised regression

Exploring underlying multivariate characteristics of two high-dimensional datasets

Discovering characteristics of Post-Treatment Recurrent High-Grade Gliomas 1) identifies distinct gene clusters corresponding to the immunohistochemical stains and 2) maximizes multivariate correlation between two datasets (imaging and genetics)

Knowledge discovery in neuroimaging data fusion for migraine diagnosis

1. Identify changes in brain function and structure that correlate with the response to erenumab (medication)

2. Develop a domain adaptation technique that tackles the challenge of discrepancy in imaging parameters across different cohorts of patients, which causes significant difficulty in building a unified imaging-based diagnostic model