The goal of MLAI lab. is to study and develop various models applicable to various domains, such as visual recognition, natural language understanding, healthcare, financial prediction, based upon the general machine learning approaches including deep learning.
- Multi-task deep learning for avoiding negative transfer on shared representations.
- Bayesian deep learning based on approximate variational inference.
- Network structure estimation and optimization for lifelong learning.
- Zero-/Few-shot learning for unseen category prediction.
- Deep-learning based survivor detection system on UAVs.
- Active incremental learning with model uncertainty for autonomous driving.
Natural language understanding
- Deep generative model based controllable text generation.
- Personalized conversation model using memory-augmented continual learning.
- Explainable AI. Uncertainty and attention mechanism based reliable prediction research.
- Physiological symptom prediction models in intensive care unit and ward environment.
- Deep gaussian process and variational approach based machine learning algorithms.
- Deep probabilistic models for algorithmic stock trading, real estate price prediction.
Scalable / Efficient Deep Networks
Devise efficient and scalable training methods for large-scale deep neural networks.
CGES: A structured sparsity regularizer that combines the (1,2)-norm (exclusive sparsity) with the (2,1)-norm (group sparsity).
SplitNet: Network split that allows each subnetwork to run completely independently on multiple GPUs.
Continual Learning / Variational Deep Learning
DEN : Dynamically expandable network partially retrains the existing network and add in only the necessary number of neurons.
DropMax : A stochastic attention method to facilitate classification to focus on fine-grained problem per instance.
Simultaneous Object/Scene Recognition and Learning from Driving Videos
Detect important objects from driving videos and incrementally train the detector with these newly detected objects.
Adpative Detection/Tracking Decision Model for Survivor Detection System
Propose a model that can automatically detect survivors in disaster scences captured from UAV and agent learns to determines whether to detect or track at each frame using a reinforcement learning.
Natural Language Understanding
Personalized Conversational Agent
Develop a personalized dialogue generation model and have been recently working on memory-augmented approach.
General/Conditional Text Generation
Study generic or conditional text generation method based on deep generative models such as GAN and VAE.
Special Research on Intensive Care Unit (ICU) Environment of Medical Hospital
Prevent economic/time loss by predicting the illness of the patient in advance.
Explainable and Reliable Medical Intelligence
Develop uncertainty and attention based deep learning models to efficiently predict and interpret causality.
Deep Learning for Forecasting Real Estate Price
Forecast real estate price considering with various factors such as commercial and geographical properties.