(Evelyn) Nahyeon Kim
Hi 😄
I am currently pursuing a PhD in Computer Science at UCLA, where I work with Prof. M. Khalid Jawed and Prof. Wei Wang on Vision Language Action (VLA) models for autonomous navigation in agricultural environments.
Before starting my PhD, I completed my Master's degree in Mechanical Engineering at UCLA in 2024, focusing on Neural Radiance Fields (NeRF)
under the guidance of Prof. M. Khalid Jawed at the Structures-Computer Interaction Lab.
I earned my Bachelor's degree in Mechanical Engineering from Seoul National University in 2023.
I am passionate about advancing AI research by integrating knowledge from diverse fields to create innovative solutions.
My primary research interest lies in developing integrated AI systems that combine vision, language, and action,
particularly for real-world applications in robotics and autonomous systems.
I am always open to discussions and collaborations. Please feel free to reach out if you would like to connect!
| Year | Organization | Role | Research Focus |
|---|---|---|---|
| 2021 | SaigeResearch Inc. | Research Intern | Optical Character Recognition (OCR) |
| 2021-2022 | Korea Institute of Science and Technology (KIST) | Research Intern | Knowledge Distillation, Neural Network Inversion |
| 2023 | VR Crew Inc. | Research Engineer | Visual Positioning System (VPS) for VR coordinate mapping |
| 2023-2024 | UCLA | M.S. Student | Neural Radiance Fields (NeRF) |
| 2024-2025 | CONNECTEVE Inc. | Research Engineer | Diffusion Models, Analytical systems for medical applications |
| 2025-Present | UCLA | Ph.D. Student | Vision-Language-Action (VLA) for autonomous driving |
PUBLICATIONS & PATENTS
[Submitted] Nahyeon Kim and Seyun Kim. “DiffusionMix: Adaptive Image Composition through Merging Two DDPM Denoising Processes.”, Pattern Recognition.
[Submitted] Nahyeon Kim and Suhyun Kim. ”KDO:Data-Free Re-training of Pruned Networks.” IEEE Access.
[Published] Nahyeon Kim. ”Advancing Neural Radiance Fields through Self-Supervised NeRF Image Selector (SNIS).” UCLA Electronic Theses and Dissertations, 2024.
[Issued] Kim, Nahyeon, 2023, Apparatus and method for performing visual localization effectively,
Korean Patent 1020230054544, filed April 26, 2023, and issued December 18, 2023.
PUBLIC PROJECTS
Self-Supervised NeRF Image Selector (SNIS)
This research focuses on optimizing the selection of camera poses to improve the NeRF model’s performance with a minimal number of images. A self-supervised learning-based NeRF Image Selector was designed to select optimal camera positions that are most advantageous for training on a target scene. The NeRF Image Selector significantly improved NeRF model training efficiency in nearly all cases by developing a novel pseudo-label. This label functions similarly to a reinforcement learning reward, enabling a self-supervised learning approach in NeRF model training. Additionally, a new framework was introduced that integrates with the Unity environment to control cameras and train NeRF models, establishing a new research methodology to advance NeRF studies.
Data-Free Retraining of Pruned Networks
Owing to concerns about privacy and licensing, there is an increasing trend of releasing large models without providing access to the training data. This makes network pruning less efficient since retraining after pruning necessitates training data. To work around the absence of the original training data, data-free pruning resorts to network inversion data generated from the model. However, since the distribution of this synthetic data differs significantly from the original, fine-tuning with synthetic data leads to limited performance improvement or even a decrease in some cases. To overcome this problem, we propose KD-only (KDO), a knowledge distillation method with only Kullback-Leibler divergence loss, as a better alternative to fine-tuning. The key intuition is that we do not trust the fidelity of the synthetic data; we use these data as the input points at which the pruned model is trained to follow the behavior of the original model. Our experiments with data-free pruning of various models demonstrate that KDO consistently outperforms fine-tuning. The effectiveness of our method remains robust even with a limited number of images or lower image quality, thereby enhancing the applicability of data-free pruning.
Comparative Analysis: Traditional Computer Vision and Deep Learning
The advent of deep learning has yielded remarkable results across various computer vision tasks. In a bid to enhance performance, several traditional computer vision techniques have been integrated into deep learning frameworks. This repository presents a comparative study examining the performance of deep neural networks when augmented with traditional computer vision algorithms: Warping, SIFT, Edge Detection, and Gabor Filters. We utilized ResNet18 and modified the first convolutional layer to adjust the number of input image channels. Our experimentation involved CIFAR10, CIFAR100, as well as high-resolution datasets such as the Oxford 102 Flower Dataset and the Large Scale Fish Dataset, aiming to enhance the performance and impact of computer vision tasks.
Spline Visualization and Scene Rendering with OpenGL
This is an OpenGL project that includes creating scenes including Implicit Surfaces and Polygonal Objects, allowing dynamic rendering and interaction with the user. Users can manipulate the camera and change scenes using mouse and keyboard controls within the visualization window. The project extends to ray tracing scenes containing objects with various materials represented by the Phong illumination model. To create polygonal objects, B-Spline and Catmull-Rom spline methods are used with coordinates of control points, which are parsed from the data file. It utilizes a total of 4 light sources and employs recursive reflection to represent light reflection.
EXPERIENCE & EDUCATION
Ph.D. Student
Structures-Computer Interaction Lab, UCLA Since Sep 2025
Supervised by Prof. M. Khalid Jawed and Prof. Wei Wang, my research focuses on Vision-Language-Action (VLA) Model for autonomous driving in farmland.
Research Engineer
CONNECTEVE.Inc Aug 2024 - Sep 2025
Supervised by Prof. Duhyun Ro, I worked on diffusion models for medical treatment applications and developed X-ray image analysis systems, including segmentation and keypoint detection algorithms, using state-of-the-art medical AI frameworks.
M.S. Student
Structures-Computer Interaction Lab, UCLA Sep 2023 - Jun 2024
Supervised by Prof. M. Khalid Jawed, my research focuses on Neural Radiance Fields (NeRF) and self-supervised learning.
Research Engineer
VR Crew Inc. Jan 2023 - Aug 2023
Developed a framework at the core of the company's Vision Positioning System, integrating point cloud analysis, computer vision deep learning, and epipolar geometry. Led research focused on Keypoint Extractor, PnP, PnL, Point Matching, and Global Descriptor.
AI Research Internship
Korea Institute of Science and Technology (KIST) Oct 2021 ~ Aug 2022
Supervised by Prof. Suhyun Kim, In KDST (KIST Data Science Team), I researched combining Knowledge Distillation and Network Inversion for Network Pruning.
AI Research Internship
Saige Research Inc. Feb 2021 - Aug 2021
Supervised by Prof. Frank Chongwoo Park, Engaged in pattern recognition and Optical Character Recognition (OCR) research and customized various supervised OCR models with multifaceted experiments.
Software Engineer Internship
PSX Inc. Aug 2020 - Feb 2021
Developed services for trading unlisted stocks, including web development using Django and hybrid app development with React-Native.
B.S. Student
Seoul National Univeisity Mar 2018 - Feb 2023
Undergraduate Studies with a special focus on programming, simulation and mathematics.