Dong-Hwan Jang

I am Dong-Hwan Jang, a recent graduate from the Master's program in Electrical and Computer Science at Seoul National University. During my academic journey, I had the privilege of receiving guidance from Professor Bohyung Han and, post-graduation, I was fortunate to collaborate with esteemed mentors Dongyoon Han and Sangdoo Yun during my internship at NAVER. Now, I serve as a research engineer at AI Research Center, Samsung Advanced Institute of Technology (SAIT).

As a machine learning researcher, my primary goal is to minimize human inductive bias in models, enabling them to learn and perform robustly across diverse environments. My initial focus was on dynamic neural networks, tailoring network structures and operations to optimize performance. Currently, my research pivots towards the area of robust fine-tuning, leveraging linear mode connectivity to enhance model robustness and effectiveness.

Email  /  CV  /  Google Scholar  /  Twitter  /  Linkedin

Thanks to our insights in the fine-tuned weight space, fine-tuning a few models (i.e., only two) can lead to superior merged weights (closer to the center of a weight space) without merging many fine-tuned models under extensive parameter searches like Model Soup.

We introduce Merge-and-Bound (M&B), an innovative approach for Class Incremental Learning that optimizes model weights through two merging techniques: inter-task and intra-task weight merging, alongside a bounded update to prevent catastrophic forgetting. Without altering architectures or objectives, M&B integrates into existing methods, showing superior performance on CIL benchmarks against top competitors.

We propose DynOPool, a learnable pooling layer that finds the optimal scale factors and receptive fields of intermediate feature maps. It adapts feature map sizes and shapes for enhanced accuracy and efficiency in tasks like image classification and semantic segmentation. This innovative approach significantly improves deep neural network optimization with minimal computational overhead.

We present DS4C South Korea Patient, Policy, and Provincial data (DS4C-PPP dataset). The dataset contains comprehensive data that could be used for causal analysis, such as per-patient symptom onset and confirmed date, travel frequency, hospital accessibility, and 61 preventative policies enacted in South Korea.

We propose spatially-variant motion deblur network based on the implicit neural representation. A spatially-variant deblurring network takes deformed features and their offsets as inputs.

We propose a novel framework of the dynamic spatially-adaptive modulation for image dehazing. The proposed algorithm introduces a selection module that conditionally determine the necessary modulation pathways in a bottom-up manner by providing a loss function optimizing both accuracy and efficiency.

We adaptively find the appropriate number of the residual blocks according to the severity and distortion type of the input in universal image restoration task.