Research Interests
The main research goal of our group is to develop novel machine learning algorithms which tackle daunting challenges of existing machine learning frameworks.
The current deep learning algorithms/frameworks are i) dependent on a massive amount of training data, ii) hard to generalize to unseen tasks and expand the past knowledge in a time stream, iii) dependent on centralized train datasets and computing resources in a massive server facilities, iv) not easy to be deployed in a real networks such as 5G and 6G environments. To tackle these problems, our group focuses on following topics: Data-efficient lifelong learning, Decentralized & federated learning framework, Advanced meta-learning algorithms, 6G intelligent communication systems. See the details below.
Data-Efficient Lifelong Learning
Modern machine learning algorithms are i) dependent on a massive amount of training data, ii) hard to generalize to unseen tasks, and iii) difficult to learn a new concept on the learned knowledge. To tackle this issue, 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.
For one of our recent work on this topic, we built a meta-learning algorithm for few-shot classification tasks with an effective vector projection technique (TapNet, ICML 2019, [paper] [video] & NeurIPS workshop 2018 [paper]).
For an advanced topic on lifelong learning, we built an incremental meta-learning algorithm that can expand classification tasks only with a few-shot of novel categories. For this algorithm, a task-adaptive representation learning module is meta-trained to extract novel features of novel categories with only a few labeled data samples (XtarNet, ICML 2020, [paper], see Fig. 1-1).
For a recent topic in computer vision, which is the class-incremental learning (FSCIL) task, we built a method that reserves the representation space in-between current classes to facilitate the acceptance of novel classes in the future (MICS, WACV 2024 [paper], see Fig. 1-2)
Figure 1-1. XtarNet with task-adaptive representation (TAR)
Figure 1-2. MICS with midpoint interpolation
Decentralized Learning & Federated Learning
Almost all state-of-the art learning algorithms are centralized in view of both data and model. However, the real environment where data samples are collected and machine learning models are trained for that data, we have to consider decentralization of neural network models, training data samples, computing/communication resources, and other related factors. Therefore, our group tries to develop decentralized learning frameworks that can collect&learn big- or small-data by utilizing distributed computing/communication/storage resources in hierarchical&clustered networks. To this end, we are studying Federated learning, Decentralized learning, Distributed systems and any related fields.
As our prior work, the relationship between storage, communication and security capacity of clustered distributed system are analyzed (two papers are published in IEEE Trans. Inf. Theory 2019 (contributed as a co-author).
The current research interest of our group is to build a federated learning framework deployed in future networks with the following functionalities: learning widely non-IID. datasets/tasks, utilizing distributed computing/storage resources, and analyzing big- or small data for both generalization & personalization (see Fig. 2).
As one of the recent results, we suggested a conjunction of meta-learning and federated learning to achieve highly-improved generalization capability in a strong non-IID case, which leads to substantial gains in personalized federated learning benchmarks (MetaVers, WACV 2024 [paper]).
Figure 2. Federated learning framework deployed in future network (e.g., 6G wireless system)
Advanced Topics for Generalization: Reinforcement Learning, Domain Generalization, Computer Vision, Semi-Supervision etc.
Reinforcement learning (RL) brings the great success in various applications including games, control systems, self-driving systems, etc. Our hope is that an intelligent system handles a wide range of related tasks even though they are unseen before. We should pursue strong robustness of RL algorithms even in real environments with unexpected events (such as accidents in an automated driving system). To tackle these issues, we are highly motivated to develop RL algorithms with strong generalization performance. One of promising directions is the Meta reinforcement learning.
Human can generalize the consistent knowledge on 'cat' on varying styles (domains) such as photo, sketch, cartoon, etc. However, modern deep models have difficulties in achieving generalization across domains. To tackle this gap, our group focuses on developing a Domain generalization method across widely varying domains. Our recent work aims to separate category-classifying embedding and the domain-classifying embedding to obtain domain-invariant features (POEM, AAAI 2023 [paper]).
Advanced computer vision applications are timely target domains of meta-learning algorithms. Our group is studying few-shot semantic segmentation methods which can generalize the learned segmentation capability to unseen object with a few labeled data samples ([arXiv paper link], see Fig. 4).
Although meta-learning can bring a strong generalization on novel tasks, a large amount of labeled data samples are required when we meta-train our model to acquire the strong generalization capability. For some practical cases, however, even the labeled data for meta-training stage cannot be prepared. For that case, we have to build semi-supervised or unsupervised learning of meta-learners to achieve the generalization capability. Our related prior work considers semi-supervised few-shot learning techniques ([arXiv paper link]).
Figure 3. Meta-learning for domain generalization
Figure 4. TAFT: Few-shot semantic segmentation algorithm
6G Intelligent Communication Systems
Almost all machine learning researchers are dedicated to develop high-performance deep learning algorithms for various applications. However, real applications are not in laboratories. The intelligent algorithms should be extended to edges, nodes or smart devices on our hands. The network infrastructures that connect massive servers and wireless devices are dependent on wireless communication systems such as 5G and beyond. 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, Machine-learning-based channel coding algorithms, Intelligent resource block allocation (RiSi, WoWMoM 2024).