MMAI Lab | Ajou University

Figure 1 from Efficient Test-time Adaptation illustrating candidate verification and divergence-shift based test-time adaptation.

NeurIPS

Computer Vision and Learning Algorithms

Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts

Seungmin Oh, Seunghun Kang, Jongbin Ryu

Annual Conference on Neural Information Processing Systems (NeurIPS)2026-12-06

Figure 1 from Differentiable Bit-Widths showing SVD-based co-optimization of pruning and quantization for ultra-efficient LLM compression.

NeurIPS

Efficient Learning for LLMs

Differentiable Bit-Widths: Co-optimizing Pruning and Quantization via SVD for Ultra-Efficient LLM Compression

Hankyul Kang, Jongbin Ryu

Annual Conference on Neural Information Processing Systems (NeurIPS)2026-12-06

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.

BMVC

Computer Vision and Learning Algorithms

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 showing FlowVLA: multimodal RGB, language, and proprioception inputs feed a VLA transformer backbone that decodes actions, with training-only dual-level supervision from reference-anchored 3D point trajectories at point and TCP flow levels.

CoRL

Robot Learning

FlowVLA: Enabling World Dynamics Understanding through Pluggable 3D Flow Learning

Junhyeong Go, Mingyu Jung, Woobin Im, Jongbin Ryu

Conference on Robot Learning (CoRL)2026-11-09

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.

EMNLP

Efficient Learning for LLMs

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.

EMNLP

Efficient Learning for LLMs

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.

EMNLP

Efficient Learning for LLMs

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 showing the pipeline: charge and discharge curves from a single cycle are transformed by a continuous wavelet transform, compressed across scales to remove redundancy, and fed to a deep neural network that predicts the remaining useful life capacity curve.

Applied Energy

Industrial and Medical AI

One-Cycle Based Battery Remaining Useful Life Prediction by Continuous Wavelet Scale Compression with Deep Neural Networks

Sungeun Kim, Mingyu Jung, Yuri Kim, Jin-Sung Park, Jongbin Ryu

Applied Energy2026-09-24

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

CVPR

Computer Vision and Learning Algorithms

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.

WACV

Computer Vision and Learning Algorithms

Channel Propagation Networks for Refreshable Vision Transformer

Junhyeong Go, 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.

WACV

Computer Vision and Learning Algorithms

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.

BSPC

Industrial and Medical AI

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

Inhyuk Park, Won Hwa Kim, Jongbin Ryu

Biomedical Signal 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.

ACCV

Industrial and Medical AI

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.

ACCV

Computer Vision and Learning Algorithms

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.

BMVC

Computer Vision and Learning Algorithms

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.

ESWA

Computer Vision and Learning Algorithms

Spatial Bias for Attention-free Non-local Neural Networks

Junhyeong Go, Jongbin 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.

CMPB

Industrial and Medical AI

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.

Sci Rep

Industrial and Medical AI

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.

MedIA

Industrial and Medical AI

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.

CIBM

Industrial and Medical AI

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.

IJS

Industrial and Medical 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 showing the proposed learning method: a backbone feeds N lightweight heads whose features are pulled toward the target by cross-entropy loss while a decorrelation loss pushes them apart from an anchor point, yielding less-correlated clusters after training.

ICCV

Computer Vision and Learning Algorithms

Gramian Attention Heads are Strong yet Efficient Vision Learners

Jongbin Ryu*, Dongyoon Han*, Jongwoo Lim (*co-first author)

International Conference on Computer Vision (ICCV)2023-10-02

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

ICCVW

Industrial and Medical AI

Robust Asymmetric Loss for Multi-Label Long-Tailed Learning

Wongi Park, Inhyuk Park, Sungeun Kim, Jongbin Ryu

International Conference on Computer Vision Workshops (ICCVW)2023-08-10

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.

JDI

Industrial and Medical AI

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.

PRL

Computer Vision and Learning Algorithms

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.

TPAMI

Computer Vision and Learning Algorithms

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.

Neural Networks

Computer Vision and Learning Algorithms

Unsupervised Feature Learning for Self-Tuning Neural Networks

Jongbin Ryu, Ming-Hsuan Yang, Jongwoo Lim

Neural Networks2021-01-01

Overview of Generalized Convolutional Forest Networks: CNN features feed random forests, whose probability mass function guides triplet sampling of anchor, positive and negative examples, and the triplet loss updates the CNN.

ICLR

Computer Vision and Learning Algorithms

Generalized Convolutional Forest Networks for Domain Generalization and Visual Recognition

Jongbin Ryu, Gitaek Kwon, Ming-Hsuan Yang, Jongwoo Lim

International Conference on Learning Representations (ICLR)2020-04-26

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.

Sci Rep

Industrial and Medical AI

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

Overview of unsupervised face domain transfer: generative face augmentation adds high-to-low resolution samples across poses, and spatial resolution adaptation then clusters high- and low-resolution faces of the same identity together.

SPL

Computer Vision and Learning Algorithms

Unsupervised Face Domain Transfer for Low-Resolution Face Recognition

Sungeun Hong, Jongbin Ryu* (*corresponding author)

IEEE Signal Processing Letters (SPL)2020-01-01

Figure 1 from OmniMVS: four fisheye input images pass through a shared 2D CNN, the unary features are warped onto spheres by spherical sweeping, and a 3D encoder-decoder with softargmin turns the cost volume into an omnidirectional depth map.

ICCV

Computer Vision and Learning Algorithms

OmniMVS: End-to-End Learning for Omnidirectional Stereo Matching

Changhee Won, Jongbin Ryu, Jongwoo Lim

International Conference on Computer Vision (ICCV)2019-10-27

Figure 4 from SweepNet: pairs of fisheye images are warped onto each virtual sphere, SweepNet computes pairwise matching costs, and the averaged cost volume is refined by cost aggregation before the depth is taken by argmin.

ICRA

Computer Vision and Learning Algorithms

SweepNet: Wide-baseline Omnidirectional Depth Estimation

Changhee Won, Jongbin Ryu, Jongwoo Lim

IEEE International Conference on Robotics and Automation (ICRA)2019-05-20

Figure 1 from the DFT pooling paper comparing feature maps at the last CNN layers: conventional layouts without and with average pooling, and the proposed DFT magnitude pooling that keeps low-frequency DFT coefficients.

ECCV

Computer Vision and Learning Algorithms

DFT-based Transformation Invariant Pooling Layer for Visual Classification

Jongbin Ryu, Ming-Hsuan Yang, Jongwoo Lim

European Conference on Computer Vision (ECCV)2018-09-08

Publication