[Research Area]

Reasoning and Interpretation of Neural Networks

• Counterfactual Reasoning via Generative Models

• Interpretation of Neural Manifold Distributions

• Reinforcement Learning towards General Intelligence

• AI Security based on High-Level Reasoning

Our research group is deeply engaged in elucidating the reasoning and interpretation of neural networks. Our objective is to demystify the logic and decision-making processes embedded within these systems, thus enhancing their transparency and interpretability. A primary focus of our work centers on counterfactual reasoning facilitated by generative models. By utilizing these models to simulate alternative scenarios, we enable a more profound understanding of model decisions through counterfactual analysis. Additionally, we investigate the neural manifold distributions, scrutinizing the geometric properties of data representations within neural networks to derive insights into their learning and information processing mechanisms.

Our endeavors extend to reinforcement learning, with the ultimate aim of advancing towards general intelligence by developing algorithms capable of learning and adapting across a broad spectrum of real-world data. Moreover, we are dedicated to enhancing AI security, concentrating on the formulation of security measures predicated on high-level reasoning. This is to safeguard against adversarial attacks and uphold the integrity of AI applications. Through these comprehensive efforts, our group contributes significantly to the deeper understanding and advancement of artificial intelligence, thus paving the way for innovative developments in the field.

Multi-Modality Learning for Cross Domain/Image Mapping

• Self-Supervised Representation Learning

• Correspondence / Relation Mapping of Neural Representations

• Effective Domain Adaptation via Generative Manifold Spaces

• Applications to Medical (e.g., multi-phase and multi-modal images) and Vision/Texts

Our research group is committed to advancing the field of multi-modality learning, with a particular emphasis on cross-domain and image mapping. We investigate how diverse data types can complement and enhance one another's analysis, a critical aspect of our work. Concurrently, we are deeply invested in self-supervised representation learning, which empowers models to acquire valuable data representations independently of labeled datasets. This approach marks a significant stride toward the realization of more autonomous AI systems. Furthermore, our research explores the correspondence and relation mapping within neural representations to elucidate the interconnectedness of different data elements within AI models, thereby improving our capacity to interpret the complex decisions made by neural networks.

A considerable portion of our efforts is also dedicated to mastering effective domain adaptation through generative manifold spaces. In this area, we formulate methods that facilitate the adaptation of AI models to novel, previously unseen domains by comprehensively understanding the data's underlying structure. Such advancements are pivotal for applications necessitating consistent model performance across varied settings without the need for extensive retraining.

The implications of our research span multiple fields, most notably in medical imaging. Here, we leverage our insights for multi-phase and multi-modal image registration and the integration of vision and text, achieving a holistic understanding across these modalities. Our objective is to forge AI technologies that are not only more versatile and efficient but also capable of integrating smoothly across varied domains and data types. This endeavor has the potential to revolutionize the application of AI in medicine, content interpretation, and beyond, significantly impacting how AI technology is utilized.

3D Vision and Medical Imaging

• Visualization of 3D Data and Image Processing

• Image Registration and 3D Modeling

• Depth-Map Enhancement and Pose Estimation

• Volume Classification, Segmentation, and Registration for Precision Medicine

Our research group excels in the advanced study of three-dimensional data and image processing, with a particular emphasis on applications that span from visual computing to precision medicine. Our work includes groundbreaking efforts in visualizing 3D data, where we develop sophisticated techniques to enhance the interpretation and analysis of complex spatial datasets. This is closely aligned with our research in image registration and 3D modeling, through which we aim to achieve precise alignment and modeling of 3D images for various applications, enhancing the accuracy of computer-assisted analysis and interpretation.

Additionally, our team is focused on improving depth-map enhancement and pose estimation. These areas are critical for developing more intuitive and interactive 3D environments, as well as for applications in augmented reality and robotics, where understanding the spatial orientation and position of objects is paramount.

A significant portion of our research is also dedicated to volume classification, segmentation, and registration within the context of precision medicine. Here, we leverage advanced 3D image processing techniques to facilitate the detailed analysis of medical scans. This work is vital for the development of personalized treatment plans, enabling more precise diagnoses and targeted therapies based on the unique anatomical and physiological characteristics of individual patients.

By combining these research efforts, our group aims to push the boundaries of what's possible in 3D data processing and image analysis, contributing to advancements that have real-world implications, especially in enhancing medical diagnostics and treatments through precision medicine.

Funded Projects:

• IITP (Institute of Information & Communications Technology Planning & Evaluation), South Korea (2024~2027).

• KEIT (Korea Evaluation Institute of Industrial Technology), South Korea (2023~2027).

• IITP (Institute of Information & Communications Technology Planning & Evaluation), South Korea (2022~2029).

• NRF (National Research Foundation of Korea), South Korea (2022~2024).

• INFINITT Healthcare Co., Ltd., South Korea (2022~2023).

• NRF (National Research Foundation of Korea), South Korea (2022~2023).

•  IITP (Institute of Information & Communications Technology Planning & Evaluation), South Korea (2021~2026).

• HealthHub Inc., South Korea (2021).

• ETRI (Electronics and Telecommunications Research Institute), South Korea (2021).