Doyeol Baek

Junsuk Choe is an associate professor in the Department of Computer Science and Engineering at Sogang University. His research focuses on multimodal foundation models, including large vision–language and video–language models, and how visual, linguistic, and temporal information is integrated and generalized. He also develops methods for controlling and adapting large-scale foundation models, focusing on continual learning, machine unlearning, and knowledge editing. A central goal of his research is to improve the reliability and efficiency of multimodal and language models at inference time, and he investigates model-level approaches to mitigating hallucinations, improving robustness and controllability, as well as inference-time optimization techniques. Additionally, he studies semi-supervised and weakly supervised learning for training and adapting vision and multimodal models with reduced reliance on large-scale manual annotations.

서강대학교 컴퓨터공학과 최준석 교수의 연구실입니다. 연구는 대규모 비전-언어 및 비디오-언어 모델을 포함한 멀티모달 파운데이션 모델, 그리고 시각, 언어, 시간 정보가 어떻게 통합되고 일반화되는지에 초점을 맞춥니다. 또한 지속 학습, 기계 언러닝, 지식 편집에 중점을 두어 대규모 파운데이션 모델을 제어하고 적응시키는 방법을 개발합니다. 연구의 핵심 목표는 추론 시 멀티모달 및 언어 모델의 신뢰성과 효율성을 향상시키는 것이며, 환각 완화, 견고성 및 제어 가능성 향상, 그리고 추론 시간 최적화 기술에 대한 모델 수준 접근 방식을 연구합니다. 또한 대규모 수동 주석에 대한 의존도를 줄여 비전 및 멀티모달 모델을 훈련하고 적응시키기 위한 준지도 학습 및 약지도 학습을 연구합니다.

Training-Free Uncertainty-guided Logit Adjustment for Few-Shot Class-Incremental Learning

Enhancing Multi-Image Understanding through Delimiter Token Scaling

What "Not" to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging

OVS Meets Continual Learning: Towards Sustainable Open-Vocabulary Segmentation

Improving Fisher Information Estimation and Efficiency for LoRA-based LLM Unlearning

LMLT : Low-to-high Multi-Level Vision Transformer for Lightweight Image Super-Resolution

Small Object Matters in Weakly Supervised Object Localization

NegMerge: Sign-Consensual Weight Merging for Machine Unlearning

VHOIP: Video-based Human-Object Interaction recognition with CLIP Prior knowledge

Fog-Free Training for Foggy Scene Understanding

Curriculum Learning with Class-label Composition for Weakly Supervised Semantic Segmentation

ConVis: Contrastive Decoding with Hallucination Visualization for Mitigating Hallucinations in Multimodal Large Language Models

Improving ViT Interpretability with Patch-level Mask Prediction

Discovering an Inference Recipe for Weakly-Supervised Object Localization

Weakly-supervised Incremental learning for Semantic segmentation with Class Hierarchy

Weakly Supervised Semantic Segmentation for Driving Scenes

Small Objects Matters in Weakly-supervised Semantic Segmentation

Neglected Free Lunch - Learning Image Classifiers Using Annotation Byproducts

Entropy Regularization for Weakly Supervised Object Localization

Evaluation for Weakly Supervised Object Localization: Protocol, Metrics, and Datasets

Weakly Supervised Semantic Segmentation using Out-of-Distribution Data

Attention-based Dropout Layer for Weakly Supervised Single Object Localization and Semantic Segmentation

Keep CALM and Improve Visual Feature Attribution

Normalization Matters in Weakly Supervised Object Localization

Contrastive Attention Maps for Self-supervised Co-localization

Rethinking Spatial Dimensions of Vision Transformers

Region-based Dropout with Attention Prior for Weakly Supervised Object Localization

Re-labeling ImageNet: from Single to Multi-Labels, from Global to Localized Labels

Evaluating Weakly Supervised Object Localization Methods Right

CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features

Attention-based Dropout Layer for Weakly Supervised Object Localization

An Empirical Evaluation on Robustness and Uncertainty of Regularization Methods

Robust approach to inverse lighting using RGB-D images

Face generation for low-shot learning using generative adversarial networks

Cutmix: Regularization strategy to train strong classifiers with localizable features

Rethinking spatial dimensions of vision transformers

Region-based dropout with attention prior for weakly supervised object localization

Contrastive Attention Maps for Self-Supervised Co-Localization

Entropy regularization for weakly supervised object localization

LMLT: Low-to-high Multi-Level Vision Transformer for Lightweight Image Super-Resolution

Weakly-Supervised Incremental Learning for Semantic Segmentation with Class Hierarchy

What “Not” to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging

OVS Meets Continual Learning: Towards Scalable Open-Vocabulary Segmentation

A Comprehensive Overhaul of Parameter-Efficient Knowledge Unlearning for LLMs

Neglected Free Lunch–Learning Image Classifiers Using Annotation Byproducts

Junsuk Choe

Associate Professor

Dongyoon Han

Senior Research Scientist and Team Leader at NAVER AI Lab, Adjunct Professor at KAIST Graduate School of AI

Junsuk Choe's Research Group

교수님 소개

최준석

연구실 소개

연구실 키워드

알럼나이가 간 회사

연구실 지원 방법

대우 조건

연도별 논문 인용수

저널 & 학회 논문 출판 수

연구실 논문 목록

Training-Free Uncertainty-guided Logit Adjustment for Few-Shot Class-Incremental Learning

Enhancing Multi-Image Understanding through Delimiter Token Scaling

What "Not" to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging

연구실 보유 특허

대학별 입시끈

우리 연구실 모집

교수님 소개

최준석

연구실 키워드

알럼나이가 간 회사

연구실 지원 방법

대우 조건

연도별 논문 인용수

저널 & 학회 논문 출판 수

연구실 논문 목록

Training-Free Uncertainty-guided Logit Adjustment for Few-Shot Class-Incremental Learning

Enhancing Multi-Image Understanding through Delimiter Token Scaling

What "Not" to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging

연구실 보유 특허

최근 3년간 연구실적 높은 연구실

대학별 입시끈

우리 연구실 모집