Main research areas & tools

  • Deep learning / Machine learning
  • Statistical signal processing
  • Information theory
  • Many areas that deal with mathematics and computation
  • Main tools: Python/NumPy, C++/CUDA, Keras, Tensorflow, Pytorch

Core machine intelligence algorithms

Neural Network based Denoising / Estimation

We apply deep neural networks to the classical denoising and estimation problems in signal processing. Under the universal denoising setting, only with the assumption of known noise model, the neural network model effectively aggregates the information from similar contexts and can attain much better performance than the original method proposed in the information theory community.

This project is supported by NRF Young Researcher Program (한국연구재단 신진연구) (2016.6~2019.5, 300 Million KRW).

Multi-task / Lifelong Reinforcement Learning

We develop new class of AI algorithms that can adapt fast and continually learn from interactive environments, such as those in autonomous digital companion like Amazon Echo. Particularly, we focus on multi-task reinforcement learning algorithm, an RL algorithm that can generalize fast to the tasks that are similar to previously learned tasks. We plan to apply ideas from universal source coding, memory-based learning and interpretable models.

This project is supported by MSIP-IITP Flagship Project on AI (주관:KAIST) (2016.12~2020.12, 750 Million KRW).

Deep Learning based External knowledge based reasoning

We tackle machine comprehension problem, such as given in Stanford Question Answering Dataset (SQuAD), with an angle of utilizing external knowledge source for answering questions for given text. Recurrent neural networks or attention mechanisms are the main wagons for this research.

This project was supprted by Samsung Electronics (2017.4~2017.12, 80 Million KRW).

Interpretable machine learning algorithms

Interpretability is an indispensable feature needed for AI algorithms that make critical decisions such as cancer treatment recommendation or load approval rate prediction. We plan to develop a new learning method for deep learning algorithms that take the interpretability into account.

This project is supported by KIST (2018.3~2020.12, 150 Million KRW).

Data science applications

Neuroscience / Medical data analyses

We apply interpretable deep learning methods to various medical / neuroscience related data for achieving high prediction accuracy and making new new scientific contributions. We try to collaborate with many partners, such as Samsung Medical Center (Prof. Ho Yun Lee), SNU Hospital (Dr. Jangsup Moon, Prof. Jaejin Song), and SKKU IBS (Prof. Choong-Wan Woo).

Satellite data based PM2.5 level estimation

We apply deep learning / machine learning method for PM2.5 level prediction using satellite and meteorological data. This is a joint work with Prof. Yang Liu (Emory University, Environmental Health)

Non-Intrusive Load Monitoring (NILM)

This project aims to disaggregate the energy usage per device from the total energy usage (time-series) data, based on convolutional neural networks (CNN). This is a joint work with Encored Technologies, Inc.

DNA sequence denoising

We apply the recently proposed Neural DUDE to the Next Generation Sequencing DNA data. The aim is to develop flexible and universal denoiser that surpass the previous state-of-the-arts that are tailored to specific sequencing devices. This is a joint work with Prof. Sungroh Yoon (SNU).