BIG DATA LAB @ INHA UNIV

"AI For Good via Accessible, Scalable, and Reliable Intelligence"

Research TOPICS

Our current research focuses on how knowledge learned by neural networks can be continually updated, recalled, recovered, transferred, and efficiently used. In particular, our work falls into one or more of the following research topics.

AI That Learns and Remembers

Memory and Knowledge Dynamics

We study how neural networks can continually learn new knowledge without forgetting previously learned knowledge. Our long-term goal is to develop neural networks that can internalize experiences, retain them over time, and recall the learned knowledge when needed.

Our research includes continual learning, human-like recall, generative memory, knowledge consolidation, selective forgetting and unlearning, continual knowledge representation, and knowledge diagnosis and repair.

Related Publications

Efficient and Adaptive AI Models

Model and Token Efficiency

We study how large-scale AI models can use only the model components, tokens, and computations required for a given task, input, or execution environment.

Our research includes neural network compression, visual and multimodal token reduction, efficient autoregressive inference, foundation model and expert routing, specialized model generation, and on-device AI. Our goal is to make large AI models more efficient and practically accessible without significant performance degradation.

Related Publications

PGB: One-Shot Pruning for BERT via Weight Grouping and Permutation, JAIR 2026
Lossless Token Merging Even Without Fine-Tuning in Vision Transformers, ECAI 2025
Training-Free Restoration of Pruned Neural Networks, arXiv — our honorary “ICML 2022” paper
QueryNet: Querying Neural Networks for Lightweight Specialized Models, Information Sciences 2022
Pool of Experts: Realtime Querying Specialized Knowledge in Massive Neural Networks, SIGMOD 2021

Transferring Knowledge without Original Data

Knowledge Transfer and Reconstruction

Trained neural networks contain knowledge learned from their training data. We study how this knowledge can be recovered, transferred, and reused without access to the original data. This is particularly important when the training data cannot be stored or shared because of privacy, ownership, storage, or communication constraints.

Our research includes data-free knowledge distillation, model inversion, synthetic knowledge reconstruction, model-based rehearsal, data-free continual learning, and knowledge transfer across different models and modalities.