Embedded Files
Machine Intelligence and Information Theory Lab @ UNIST Graduate School of AI & Department of EE
The overarching theme of our group is establishing the theoretical foundation of cutting-edge machine learning algorithms and building principled learning algorithm for robust/reliable intelligent systems.
The overarching theme of our group is establishing the theoretical foundation of cutting-edge machine learning algorithms and building principled learning algorithm for robust/reliable intelligent systems.
We are mainly interested in learning theory to provide the fundamental theoretical understanding of machine intelligence ('MI') via information theory ('IT'). Based on theory, we further develop robust/reliable intelligent systems from the perspectives, including generalization, robustness, safety, etc. See the details below.
I. Generalization and Robustness
I. Generalization and Robustness
Modern deep AI requires sufficient generalization across different tasks and demands to be robust against out-of-distribution (e.g., long-tail problems).
Our group tries to establish principles of Generalization and Robustness of deep models across frameworks, including reinforcement learning models, deep generative models, multimodal models, vision-language and action models.
Another branch is the Reliability of deep models against unintended or adversarial disruptions, including data corruptions and model/data adversary.
Our recent works in this scope:
[AAAI'26] A Flat Minima Perspective on Understanding Augmentations and Model Robustness [arXiv]
[ICCV'25b] Understanding Flatness in Generative Models: Its Role and Benefits [paper] (collaboration with LAIT Lab at UNIST led by Prof. Jaejun Yoo)
[ICCV'25a] Can One Modality Model Synergize Training of Other Modality Models? [paper]
[ICLR'25] Flat Reward in Policy Parameter Space Implies Robust Reinforcement Learning [paper]
[AAAI'23] POEM: Polarization of Embeddings for Domain-Invariant Representations [paper]
Figure I-1. [ICLR'25] Robust path finding policy (right) over the baseline (left)
Figure I-2. [ICCV'25] Diffusion model with flat minima enable diverse data generation
II. Data-Efficient Adaptation and Continual Learning
II. Data-Efficient Adaptation and Continual Learning
Our group focuses on lifelong learning algorithms that can expand learned knowledge with a few shots of training data samples while enhancing past knowledge. To this end, we are researching Meta-learning, Continual learning, Incremental learning, Few-shot learning, and any related topics.
Our recent works in this scope:
[AISTATS'24] XB-MAML: Learning Expandable Basis Parameters for Effective Meta-Learning with Wide Task Coverage [paper]
[WACV'24a] MICS: Midpoint Interpolation to Learn Compact and Separated Representations for Few-Shot Class-Incremental Learning [paper]
[ICML'20] XtarNet: Learning to Extract Task-Adaptive Representation for Incremental Few-Shot Learning [paper]
[ICML'19] TapNet: Neural Network Augmented with Task-Adaptive Projection for Few-Shot Learning [paper]
Figure II-1. [ICML'20] XtarNet with task-adaptive representation (TAR)
Figure II-2. [WACV'24] MICS with midpoint interpolation
III. Decentralized Learning and Safeness/Privacy
III. Decentralized Learning and Safeness/Privacy
Our group develops decentralized learning frameworks utilizing distributed computing/communication/storage resources in complicated networks. To this end, we are studying Federated learning (FL), Decentralized learning, Distributed systems, and any related fields.
For another branch, we aim to understand privacy and security issues raised from the perspectives of data/knowledge/model. This includes our research topics of Unlearning, Inversion-attack, and theoretical analysis on the privacy of modern deep models.
Our recent works in this scope:
[CVPR'26] Pose-guided Enriched Feature Learning for Federated-by-camera Person Re-identification (collaboration with VIP Lab at UNIST led by Prof. Jae-Young Sim)
[PRL'25] Benchmarking Federated Learning for Semantic Datasets: Federated Scene Graph Generation [paper] (collaboration with Dr. Jiyoun Lim at ETRI)
[ICML'24] Rethinking the Flat Minima Searching in Federated Learning [paper]
[WACV'24b] MetaVers: Meta-Learned Versatile Representations for Personalized Federated Learning [paper]
Figure III-1. FL framework deployed in future networks
Figure III-2. [ICML'24] FedGF with generalized FL
IV. Intelligent Communication Systems & Information Theory
IV. Intelligent Communication Systems & Information Theory
To support intelligent services and applications such as autonomous driving, real-time image applications/language processing, etc., wireless communication systems also should be INTELLIGENT. Our group's principal investigator studied information theory, 5G wireless communications, channel coding algorithms. In this line, we are trying to develop intelligent communication algorithms for realizing 6G intelligent networks.
To be specific, we are studying RAN-agnostic intelligent communications, Semantic communications, ML-based RAN/Protocols, etc.
Our recent works in this scope:
V. Next-Generation Learning Paradigms
V. Next-Generation Learning Paradigms
Beyond the current deep architectures and gradient-descent paradigm, our group aims to envision next-generation learning paradigms, including Quantum Machine Learning. We try to discover the unseen values of brand-new learning systems/algorithms, while seeking the synergy between the current DL regimes and novel training paradigms, such as Quantum architecture.
Our ongoing works are submitted or in-preparation.
Google Sites
Report abuse