MMAI Lab | Ajou University

Figure 1 illustrating learnable text prompts and an augmented contrastive space built from image and text encoders over transformed image-text pairs, with adaptive re-calibration reweighting the contrastive loss targets.

Computer Vision and Learning Algorithms

Vision-Language ModelsContrastive LearningPrompt Conditioning

Re-calibrated Contrastive Loss for Transformation-Aware Prompt Conditioning in Vision-Language Models

Seungmin Oh, Seunghun Kang, Jongbin Ryu

British Machine Vision Conference (BMVC)2026-11-23

Figure 1 comparing a baseline MoE quantization ensemble that mixes a poorly performing expert into the router's output against Colla-Q, which selects and ensembles collaborative experts for a better aggregated result.

Efficient Learning for LLMs

Mixture-of-ExpertEnsembleQuantization

Colla-Q: Toward Collaborative Experts in MoE Quantization via Minimax Precision Balancing

EunJu Shin, Jongbin Ryu

Conference on Empirical Methods in Natural Language Processing (EMNLP)2026-10-24

Figure 1 comparing a baseline that optimizes all student layers at once, causing cumulative error to grow with depth, against the proposed easy-to-hard layer-wise curriculum that passes curriculum features between layers to keep cumulative error low.

Efficient Learning for LLMs

Curriculum LearningLLM CompressionHardware Acceleration

Layer-wise Curriculum Learning for Efficient LLM Compression

Donggeon Lee, Dooyeon Na, Seungmin Oh, Jongbin Ryu

Conference on Empirical Methods in Natural Language Processing (EMNLP)2026-10-24

Figure 1 comparing LoRA, which keeps pruned layers frozen with low-rank adapters, against the proposed OverRep method, which trains an overcomplete parameterization that is re-parameterized into a compact form at deployment.

Efficient Learning for LLMs

Structured LLM PruningOvercomplete Re-parameterizationCapacity-knowledge asymmetry

Train Overcomplete, Deploy Compact: Scaling Recovery Capacity for Structured LLM Pruning

Seungmin Oh, Donggeon Lee, Jongbin Ryu

Conference on Empirical Methods in Natural Language Processing (EMNLP)2026-10-24

Figure 1 from the Respacing paper illustrating the forgetting curve and how recall interval affects memory retention decay.

Computer Vision and Learning Algorithms

Continual LearningSelf-Supervised LearningReplay Methods

Do Your Best and Get Enough Rest for Continual Learning

Hankyul Kang, Gregor Seifer, Donghyun Lee, Jongbin Ryu

Computer Vision and Pattern Recognition (CVPR)2025-06-01

Figure 1 comparing Channel Propagation networks (CP-Swin, CP-PiT) with baseline Swin and PiT across parameter count and top-1 accuracy.

Computer Vision and Learning Algorithms

Vision TransformerChannel PropagationOver-smoothing

Channel Propagation Networks for Refreshable Vision Transformer

Junhyung Ko, Jongbin Ryu

Winter Conference on Applications of Computer Vision (WACV)2025-01-01

Figure 1 comparing MAP-ConvNeXt and MAP-MaxViT against SOTA networks on throughput versus top-1 accuracy for small and large models.

Computer Vision and Learning Algorithms

Attention PoolingLocal PatternsFeature Representation

Enriching Local Patterns with Multi-Token Attention for Broad-Sight Neural Networks

Hankyul Kang, Jongbin Ryu

Winter Conference on Applications of Computer Vision (WACV)2025-01-01

Figure 1 from Style-KD showing source and reference retinal images used to balance the APTOS2019 training set.

Industrial and Medical AI

Class ImbalanceKnowledge DistillationMedical Image Classification

Style-KD : Class-Imbalanced Medical Image Classification via Style Knowledge Distillation

Inhyuk Park, Won Hwa Kim, Jongbin Ryu

Biomedical Signall Processing and Control2024-12-29

Figure 2 showing the generative self-supervised learning pipeline: LLM-elaborated prompts, diffusion-based image generation, and AdaIN-based reconstruction.

Industrial and Medical AI

Generative Self-Supervised LearningMedical Image ClassificationDiffusion Model

Generative Self-Supervised Learning for Medical Image Classification

Inhyuk Park, Sungeun Kim, Jongbin Ryu

Asian Conference on Computer Vision (ACCV)2024-12-01

Figure 1 illustrating the progression from block-level to branch-level connectivity in the re-parameterization network.

Computer Vision and Learning Algorithms

Network Re-parameterizationBranch-level ConnectivityModel Efficiency

Neural Substitution for Branch-level Network Re-parameterization

Seungmin Oh, Jongbin Ryu

Asian Conference on Computer Vision (ACCV)2024-12-01

Figure 1 comparing clustering results of a traditional hashing algorithm with the proposed Hyper Quantization Tree, plus the HQT training pipeline.

Computer Vision and Learning Algorithms

Unsupervised HashingImage RetrievalQuantization

Unsupervised Hashing Network with Hyper Quantization Tree

Sungeun Kim, Jongbin Ryu

British Machine Vision Conference (BMVC)2024-11-25

Figure 1 from Spatial Bias comparing inference time and top-1 accuracy across non-local network variants.

Computer Vision and Learning Algorithms

Non-local NetworksAttention-free VisionSpatial Bias

Spatial Bias for Attention-free Non-local Neural Networks

Junhyung Go, Jonngbin Ryu

Expert Systems with applications2024-10-04

Figure 2 from ADNet comparing a conventional U-Net decoder with the proposed attentional decoder and harmonic magnitude transform.

Industrial and Medical AI

Chest X-rayAttention DecoderHigh-resolution Features

Attentional Decoder Networks for Chest X-ray Image Recognition on High-resolution Features

Hankyul Kang, Namkug Kim, Jongbin Ryu

Computer Methods and Programs in Biomedicine2024-04-21

Figure 1 from the CNN and ViT analysis paper showing robustness results and corrupted medical image examples.

Industrial and Medical AI

Medical ImagingCNN and ViT AnalysisModel Interpretability

Analyzing to Discover Origins of CNNs and ViT Architectures in Medical Images

Seungmin Oh, Namkug Kim, Jongbin Ryu

Scientific Reports2024-04-16

Figure 1 showing the long-tailed distribution of clinical findings in the CXR-LT 2023 challenge dataset, highlighting newly added versus original labels.

Industrial and Medical AI

Long-tailed LearningMulti-label ClassificationChest X-ray

Towards long-tailed, multi-label disease classification from chest X-ray: Overview of the CXR-LT challenge

CVAMD participants including Jongbin Ryu

Medical Image Analysis (MedIA)2024-01-01

Figure 1 showing the FG-SSL architecture with shuffled and distorted image branches feeding a shared hierarchical block toward a Barlow Twins style cross-correlation loss.

Industrial and Medical AI

Fine-Grained ClassificationSelf-Supervised LearningJigsaw Puzzle

Fine-Grained Self-Supervised Learning with Jigsaw puzzles for medical image classification

Wongi Park, Jongbin Ryu

Computers in Biology and Medicine2024-01-01

Figure 2 from the acute small bowel obstruction study showing the proposed CT diagnosis network workflow.

Industrial and Medical AI

CT AnalysisRisk PredictionClinical AI

Deep Learning Using Computed Tomography to Identify High-Risk Patients for Acute Small Bowel Obstruction: Development and Validation of a Prediction Model: A Retrospective Cohort Study

Seungmin Oh, Jongbin Ryu, Ho-Jung Shin, Jeong Ho Song, Sang-Yong Son, Hoon Hur, Sang-Uk Han

International Journal of Surgery2023-12-01

Figure 2 from the RAL paper comparing BCE, ASL, and robust asymmetric loss probabilities on multi-label and single-label medical images.

Industrial and Medical AI

Long-tailed LearningMulti-label ClassificationRobust Loss

Robust Asymmetric Loss for Multi-Label Long-Tailed Learning

Wongi Park, Inhyuk Park, Sungeun Kim, Jongbin Ryu

ICCVW2023-08-10

Figure 2 showing the study pipeline: a DICOM input image is contrast-converted by setting window width and level, optionally converted to JPEG, then classified by a ResNet-50 backbone into pneumothorax or normal.

Industrial and Medical AI

Chest RadiographyPneumothorax DetectionMedical Image Analysis

Effect of Contrast Level and Image Format on a Deep Learning Algorithm for the Detection of Pneumothorax with Chest Radiography

Myeong Seong Yoon, Gitaek Kwon, Jaehoon Oh*, Jongbin Ryu*, Jongwoo Lim, Bo-kyeong Kang, Juncheol Lee, Dong-Kyoon Han (*corresponding author)

Journal of Digital Imaging2023-01-25

Figure 1 illustrating the dual aggregation design: multi-scale activation maps are down-channeled and combined in a dense aggregation layer for feature level aggregation, then passed to class-wise classifiers whose positive scores are aggregated at the classifier level.

Computer Vision and Learning Algorithms

Feature Pyramid NetworkMulti-Label ClassificationFeature Aggregation

Dual Aggregated Feature Pyramid Network for Multi Label Classification

Dongjoo Yun, Jongbin Ryu*, Jongwoo Lim (*corresponding author)

Pattern Recognition Letters2021-01-01

Example results showing four wide-angle fisheye camera inputs on the left and, on the right, the reconstructed panorama with its estimated inverse depth map and uncertainty map.

Computer Vision and Learning Algorithms

Omnidirectional StereoDepth EstimationUncertainty Estimation

End-to-End Learning for Omnidirectional Stereo Matching with Uncertainty Prior

Changhee Won, Jongbin Ryu, Jongwoo Lim

IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)2021-01-01

Figure 1 showing the self-tuning pipeline: a mini-batch with an anchor is encoded by a convolutional network, grouped by bagged clustering, ranked by clustering and distance to reveal violations, and used to form triplets.

Computer Vision and Learning Algorithms

Unsupervised Feature LearningSelf-Tuning NetworksRepresentation Learning

Unsupervised Feature Learning for Self-Tuning Neural Networks

Jongbin Ryu, Ming-Hsuan Yang, Jongwoo Lim

Neural Networks2021-01-01

Figure 3 showing pairs of paediatric abdominal radiographs and their class activation maps, where the heatmap overlay highlights the region the model used to detect intussusception.

Industrial and Medical AI

Abdominal RadiographyIntussusception DetectionVisual Explanation

Deep Learning Algorithms for Detecting and Visualising Intussusception on Plain Abdominal Radiography in Children: A Retrospective Multicenter Study

Gitaek Kwon*, Jongbin Ryu*, Jaehoon Oh, Jongwoo Lim, Bo-kyeong Kang, Chiwon Ahn, Junwon Bae, Dong Keon Lee (*co-first author)

Scientific Reports2020-01-01

Publication

Filter and search

Publication Archive

Re-calibrated Contrastive Loss for Transformation-Aware Prompt Conditioning in Vision-Language Models

Colla-Q: Toward Collaborative Experts in MoE Quantization via Minimax Precision Balancing

Layer-wise Curriculum Learning for Efficient LLM Compression

Train Overcomplete, Deploy Compact: Scaling Recovery Capacity for Structured LLM Pruning

Do Your Best and Get Enough Rest for Continual Learning

Channel Propagation Networks for Refreshable Vision Transformer

Enriching Local Patterns with Multi-Token Attention for Broad-Sight Neural Networks

Style-KD : Class-Imbalanced Medical Image Classification via Style Knowledge Distillation

Generative Self-Supervised Learning for Medical Image Classification

Neural Substitution for Branch-level Network Re-parameterization

Unsupervised Hashing Network with Hyper Quantization Tree

Spatial Bias for Attention-free Non-local Neural Networks

Attentional Decoder Networks for Chest X-ray Image Recognition on High-resolution Features

Analyzing to Discover Origins of CNNs and ViT Architectures in Medical Images

Towards long-tailed, multi-label disease classification from chest X-ray: Overview of the CXR-LT challenge

Fine-Grained Self-Supervised Learning with Jigsaw puzzles for medical image classification

Deep Learning Using Computed Tomography to Identify High-Risk Patients for Acute Small Bowel Obstruction: Development and Validation of a Prediction Model: A Retrospective Cohort Study

Robust Asymmetric Loss for Multi-Label Long-Tailed Learning

Effect of Contrast Level and Image Format on a Deep Learning Algorithm for the Detection of Pneumothorax with Chest Radiography

Dual Aggregated Feature Pyramid Network for Multi Label Classification

End-to-End Learning for Omnidirectional Stereo Matching with Uncertainty Prior

Unsupervised Feature Learning for Self-Tuning Neural Networks

Deep Learning Algorithms for Detecting and Visualising Intussusception on Plain Abdominal Radiography in Children: A Retrospective Multicenter Study