Author name cluster

MinJun Kim

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

7 papers

1 author row

AAAI Conference 2026 Conference Paper

LampQ: Towards Accurate Layer-wise Mixed Precision Quantization for Vision Transformers

MinJun Kim
Jaeri Lee
Jongjin Kim
Jeongin Yun
Yongmo Kwon
U Kang

How can we accurately quantize a pre-trained Vision Transformer model? Quantization algorithms compress Vision Transformers (ViTs) into low-bit formats, reducing memory and computation demands with minimal accuracy degradation. However, existing methods rely on uniform precision, ignoring the diverse sensitivity of ViT components to quantization. Metric-based Mixed Precision Quantization (MPQ) is a promising alternative, but previous MPQ methods for ViTs suffer from three major limitations: 1) coarse granularity, 2) mismatch in metric scale across component types, and 3) quantization-unaware bit allocation. In this paper, we propose LampQ (Layer-wise Mixed Precision Quantization for Vision Transformers), an accurate metric-based MPQ method for ViTs to overcome these limitations. LampQ performs layer-wise quantization to achieve both fine-grained control and efficient acceleration, incorporating a type-aware Fisher-based metric to measure sensitivity. Then, LampQ assigns bit-widths optimally through integer linear programming and further updates them iteratively. Extensive experiments show that LampQ provides the state-of-the-art performance in quantizing ViTs pre-trained on various tasks such as image classification, object detection, and zero-shot quantization.

PDF Details DOI

AAAI Conference 2026 Conference Paper

Neural Collapse-Informed Initialization with Perturbation Injection in Classification-based Metric Learning

Jinhee Park
Hee bin Yoo
MinJun Kim
Byoung-Tak Zhang
Junseok Kwon

Recent studies have revealed Neural Collapse (NC) in deep classifiers, where last-layer weights and features align into an equiangular tight frame (ETF), concentrating class information along specific embedding directions. However, conventional fine-tuning typically disregards this structure, initializing task-specific classifier heads randomly. To explicitly leverage this phenomenon, we propose a simple yet effective method for metric learning: (1) initializing the classifier head along each class’s NC direction from a pretrained model to preserve the emergent structure, and (2) injecting small isotropic Gaussian noise during finetuning to boost generalization. In addition, we provide a theoretical bound proving that our method explicitly reduces cumulative weight drift from the NC-initialization, compared to standard finetuning. This suggests that our method better preserves the pretrained model’s class-specific structure. Empirically, this structural preservation yields Recall@K gains: reduced weight drift correlates with better performance. Concurrent decreases in the Neural Collapse 1 (NC1) measure confirm that stronger intra‐class cohesion underlies these improvements. Furthermore, we validate the effectiveness of our method on class‐imbalanced benchmarks.

PDF Details DOI

YNIMG Journal 2025 Journal Article

In-vivo high-resolution χ-separation at 7T

Jiye Kim
MinJun Kim
Sooyeon Ji
Kyeongseon Min
Hwihun Jeong
Hyeong-Geol Shin
Chungseok Oh
Robert J. Fox

A recently introduced quantitative susceptibility mapping (QSM) technique, χ-separation, offers the capability to separate paramagnetic (χpara ) and diamagnetic (χdia ) susceptibility distribution within the brain. In-vivo high-resolution mapping of iron and myelin distribution, estimated by χ-separation, could provide a deeper understanding of brain substructures, assisting the investigation of their functions and alterations. This can be achieved using 7T MRI, which benefits from a high signal-to-noise ratio and susceptibility effects. However, applying χ-separation at 7T presents difficulties due to the requirement of an R2 map, coupled with issues such as high specific absorption rate (SAR), large B1 transmit field inhomogeneities, and prolonged scan time. To address these challenges, we developed a novel deep neural network, R2PRIMEnet7T, designed to convert a 7T R2* map into a 3T R2′ map. Building on this development, we present a new pipeline for χ-separation at 7T, enabling us to generate high-resolution χ-separation maps from multi-echo gradient-echo data. The proposed method is compared with alternative pipelines, such as an end-to-end network and linearly-scaled R2′, and is validated against χ-separation maps at 3T, demonstrating its accuracy. The 7T χ-separation maps generated by the proposed method exhibit similar contrasts to those from 3T, while 7T high-resolution maps offer enhanced clarity and detail. Quantitative analysis confirms that the proposed method surpasses the alternative pipelines. The proposed method results well delineate the detailed brain structures associated with iron and myelin. This new pipeline holds promise for analyzing iron and myelin concentration changes in various neurodegenerative diseases through precise structural examination.

Details DOI

YNIMG Journal 2025 Journal Article

Resolution generalization of deep learning-based dipole inversion networks for QSM

Sooyeon Ji
MinJun Kim
Jongho Lee
Hyeong-Geol Shin

Deep learning-based dipole inversion networks for quantitative susceptibility mapping (QSM) display low performance when test data resolution is different from network-trained data resolution. While several approaches were proposed to enhance the dipole inversion networks’ resolution generalizability, they modify network architecture or parameter, limiting direct application to existing pre-trained dipole inversion networks. This study presents a novel pipeline that enables pre-trained dipole inversion networks to reconstruct QSM from input local field maps of various resolutions. The developed pipeline consisted of four steps. (ⅰ) The local field map was re-sampled at multiple different spatial locations, generating multiple local field maps at network-trained resolution. (ⅱ) The re-sampled local field maps were inferred through the network, generating QSM maps. (ⅲ) These QSM maps were combined, and then (ⅳ) compensated for systematic errors, introduced by the previous re-sampling and combining process, by “dipole compensation”. The performance of the proposed pipeline was compared with two alternative pipelines using the same network: interpolating the input data to the trained resolution prior to inference (interpolation pipeline), and naïvely inferencing (naïve-input pipeline). Through qualitative and quantitative evaluations, we demonstrate that the proposed pipeline displays superior performance compared to the alternative pipelines. Specifically, when a local field map of 1 mm3 resolution was tested using QSMnet pre-trained at 1. 5 mm3 resolution, the proposed pipeline outperformed the two alternative pipelines (NRMSE: 43. 1/49. 3/56. 0, SSIM: 0. 933/0. 910/0. 920, PSNR: 47. 1/46. 0/44. 8, HFEN: 39. 9/40. 8/48. 0 for proposed/interpolation/naïve-input pipeline). This study provides a promising solution for enhancing the generalizability of pre-trained dipole inversion networks to different input data resolutions, widening their applications in clinical settings.

Details DOI

IJCAI Conference 2025 Conference Paper

Zero-shot Quantization: A Comprehensive Survey

MinJun Kim
Jaehyeon Choi
Jongkeun Lee
Wonjin Cho
U Kang

Network quantization has proven to be a powerful approach to reduce the memory and computational demands of deep learning models for deployment on resource-constrained devices. However, traditional quantization methods often rely on access to training data, which is impractical in many real-world scenarios due to privacy, security, or regulatory constraints. Zero-shot Quantization (ZSQ) emerges as a promising solution, achieving quantization without requiring any real data. In this paper, we provide a comprehensive overview of ZSQ methods and their recent advancements. First, we provide a formal definition of the ZSQ problem and highlight the key challenges. Then, we categorize the existing ZSQ methods into classes based on data generation strategies, and analyze their motivations, core ideas, and key takeaways. Lastly, we suggest future research directions to address the remaining limitations and advance the field of ZSQ. To the best of our knowledge, this paper is the first in-depth survey on ZSQ.

PDF Details DOI

AAAI Conference 2024 Conference Paper

BOK-VQA: Bilingual outside Knowledge-Based Visual Question Answering via Graph Representation Pretraining

MinJun Kim
SeungWoo Song
YouHan Lee
Haneol Jang
KyungTae Lim

The current research direction in generative models, such as the recently developed GPT4, aims to find relevant knowledge information for multimodal and multilingual inputs to provide answers. Under these research circumstances, the demand for multilingual evaluation of visual question answering (VQA) tasks, a representative task of multimodal systems, has increased. Accordingly, we propose a bilingual outside-knowledge VQA (BOK-VQA) dataset in this study that can be extended to multilingualism. The proposed data include 17K images, 17K question-answer pairs for both Korean and English and 280K instances of knowledge information related to question-answer content. We also present a framework that can effectively inject knowledge information into a VQA system by pretraining the knowledge information of BOK-VQA data in the form of graph embeddings. Finally, through in-depth analysis, we demonstrated the actual effect of the knowledge information contained in the constructed training data on VQA.

PDF Details DOI

YNIMG Journal 2023 Journal Article

Depth-wise profiles of iron and myelin in the cortex and white matter using χ-separation: A preliminary study

Subin Lee
Hyeong-Geol Shin
MinJun Kim
Jongho Lee

The in-vivo profiling of iron and myelin across cortical depths and underlying white matter has important implications for advancing knowledge about their roles in brain development and degeneration. Here, we utilize χ-separation, a recently-proposed advanced susceptibility mapping that creates positive ( χ p o s ) and negative ( χ n e g ) susceptibility maps, to generate the depth-wise profiles of χ p o s and χ n e g as surrogate biomarkers for iron and myelin, respectively. Two regional sulcal fundi of precentral and middle frontal areas are profiled and compared with findings from previous studies. The results show that the χ p o s profiles peak at superificial white matter (SWM), which is an area beneath cortical gray matter known to have the highest accumulation of iron within the cortex and white matter. On the other hand, the χ n e g profiles increase in SWM toward deeper white matter. These characteristics in the two profiles are in agreement with histological findings of iron and myelin. Furthermore, the χ n e g profiles report regional differences that agree with well-known distributions of myelin concentration. When the two profiles are compared with those of QSM and R2*, different shapes and peak locations are observed. This preliminary study offers an insight into one of the possible applications of χ-separation for exploring microstructural information of the human brain, as well as clinical applications in monitoring changes of iron and myelin in related diseases.

Details DOI