In this project, we aim to solve the short-term prediction of precipitation (rainfall) up to ~5 days in the future. Weather agencies currently use physics-based numerical weather prediction (NWP) models to solve this problem, but these may contain errors. To correct these errors, we develop transformer-based models to post-process the output of existing NWP models to generate better precipitation predictions. We use 10+ years of spatio-temporal data over the Korean peninsula from weather observations and NWP model outputs to train our models. Our key research interests include designing the learning framework and model architecture, input variable selection, and deep ensembling from multiple NWP sources.

This project aims to develop a reinforcement learning (RL) framework for supporting military decision-making processes in combat situations. To achieve this goal, our lab focuses on 1. Hierarchical RL for multi-level decision-making, 2. Model-based RL with diffusion models for planning, 3. Extending these algorithms to multi-agent cases.

This project aims to develop an algorithm for adaptive and lightweight edge-collaborative research. We investigate and analyze S.O.T.A. federated learning algorithms which are robust to heterogeneous data. Besides, we verify our novel federated learning algorithm in a real edge-collaborative scenario where Jetson nano devices are used.

In this year, we aim to develop efficient LLM Inference algorithm (e.g., speculative decoding). Below publications [1,2] are the results of last year's project which were about knowledge distillation and pruning. [1] Revisiting Intermediate Layer Distillation for Compressing Language Models: An Overfitting Perspective (Findings of EACL2023) [2] NASH: A Simple Unified Framework of Structured Pruning for Accelerating Encoder-Decoder Language Models (Findings of EMNLP2023)

This project is aimed at developing a collaborative learning framework for multi-agent systems. As part of a larger research initiative, our laboratory focuses on federated learning within the context of data-heterogeneity among multiple agents. We aim to develop both general and personalized federated learning algorithms that surpass existing state-of-the-art solutions.

Goal-oriented Reinforcement Learning (GoRL) is a framework where agents are expected to complete similar tasks and learn to reach multiple goals. This project, joint with four other labs, focuses on developing a GoRL framework for multi-contact robot motion planning. Our lab focuses on developing theoretical foundations of the GoRL framework, focusing on 1. theoretical analyses of RL problems (bandits, MDPs) with structural assumptions, and 2. mathematically well-founded approach to various realistic scenarios (e.g., networked bandits).

This project aims to create meta-learning algorithms for the composition of new human voices. To accurately synthesize new voices, we might use multi-modal information such as bio-streamlines from sensors, visual signals and audio data. In this project, our team plays a crucial role in developing meta-learning algorithms for low-task regimes. So far, meta-learning frameworks that rapidly adapt a novel tasks to guarantee generalization performance have assumed a large number of training tasks. Recent studies, however, reveal that meta-learning algorithms also impairs their generalization when the limited number of tasks is given. Consequently, we develop a new regularization techniques to preserve generalization performance of meta-learning algorithms.

This project aims to solve the short-term prediction of particulate matter up to ~3 days by using artificial intelligence. Traditional approaches used physics-based numerical models, such as CMAQ, to predict particulate matter in Korea. However, since they showed lots of errors, we are developing AI models to correct the predictions from existing CMAQ models. After making a good benchmark, we plan to enhance performance and design some novel model architectures or methodologies.

This project aims at KEPCO to form a cluster of five schools including KAIST to propose XAI-based smart energy platform technology. In our lab, we study to solve bottlenecks caused by gradient sharing that occurs when training deep learning models with large amounts of data. To solve this problem, we compress the gradient using methods such as Random Projection and Sequential Transmitting.

In this project, the main goal is to develop an efficient algorithm for the deployment of Edge AI. We survey the related model compression algorithms and develop a unified algorithm considering the regularization technique, pruning, knowledge distillation, and neural architecture search methods.

Development of AI models for predicting various properties of tire compounds

Development of AI models for recommending tire recipes

The main goal of this project is to develop an efficient automated hyperparameter search algorithm for ML algorithms. Recent advances in ML algorithms lead to better generalization ability while they require expert knowledge for training and tuning such models. Automated hyperparameter search algorithms can alleviate this issue, but most literature suffers from high resource budgets for the search phase. We analyze the mechanism of such algorithms and develop a novel algorithm enabling rapid search with fewer search rounds with Bayesian optimization approach.

We often unavailable to obtain enough labeled data for our target task(ex. Image Classification) in most real-world cases. Some recent works address the data shortage by proposing un-, semi-, or self- supervised learning methods that utilize unlabeled data to improve the performance in target task without or with only a few amounts of labeled data. In this project, we extend those methods to build a data and training efficient algorithm to cope with various real-world constraints. We start by vision data and then consider sequential data.

Even though we can get a wide range of data while developing, producing and testing the semiconductors, it is hard to analyze such unstructured data. However, if we identify the relationship between data and develop a model for them, we can expect to improve its productivity. Since the deep learning algorithms have shown the promising ability of modeling complex data, this project aims to analyze semiconductor process data utilizing various algorithms.

The goal of this project is to develop AI models for reliable prediction & recommending tire recipes.

The global pandemic of COVID-19 that has continued from 2019 to the present indicates the importance of early diagnosis of infectious diseases. If we can detect in advance that a disease has a high risk of spreading significantly, it will not only be able to specify information that needs to be searched, but also prevent the spread of infectious diseases by appropriate early responses. In this project, we develop an infectious disease monitoring/surveillance model, so that important information is analyzed to quickly grasp the risk in the early stages of infectious diseases. Also, by integrating the monitoring model and the spreading model, we enable complex analysis of infectious diseases.

The goal of the project is to create a Neural Architecture Search (NAS) algorithm that can detect communication network failures. We aim to address two issues: (1) Topological dataset: practically all NAS algorithms have concentrated on well-known dataset like image and language, and these algorithms are unable to use topological information effectively. (2) Correlation between network layers: because a communication system is made up of numerous layers that are very tightly linked, all data from each layer must be treated carefully.

This project aims to develop fast adaptive models in online learning environments. In the online learning environments with partial feedback, it is required to consider the exploration and exploitation tradeoff for decision-making. To control the tradeoff, we focus on bandit strategies. Bandit strategies are the techniques to optimize decision-making at each round based on historical data. By utilizing the bandit strategies, we suggest algorithms to learn model parameters accurately while maximizing reward. Our suggested models can be applied to many machine learning problems such as recommender systems and hyper-parameter tuning.

In the project, our focus is to discover new types of molecules using Quantum Machine Learning techniques. Current generative models , such as diffusion models, can help create new types of molecules, but often, the molecule created doesn't satisfy desired chemical properties. Our research goal is to leverage quantum machine learning tools to develop a new type of diffusion model that is better at creating molecules with good chemical properties.