Reinforcement Learning (in progress)

View all publications


Radar System based on Deep Learning
Our group is working on radar system design based on in-depth learning. We introduce in-depth study of radar system design to identify the location, velocity and label of objects. Deep learning-based radar systems can improve performance in both directions. First, it is robust to strong interference signals such as scattering of signals from buildings and ground. Second, you can safely distinguish two objects with similar radar cross sections, such as unmanned aerial vehicles and birds.
Memory-Augmented Neural Networks
When we must both preserve information for later use and perform some computations for immediate use, separating memory storage from computation is a feasible choice. The architectures that neural networks are connected with external memory called MANNs (Memory-Augmented Neural Networks). We are developing and investigating MANN structures and trying to apply them to natural language processing tasks.
Storage Device Diagnosis via Deep Learning
Storage device can be shut down without any proactive warning. Our goal is to diagnose the state of storage device by estimating the probability of failure from its complex S.M.A.R.T. signals. Since the signal is high-dimensional, sequential, noisy, and unbalanced, we are trying to handle this by deep learning with some techniques.
Hot / Cold Identification via Deep Learning
Defining hot data as the data likely to be rewritten, the durability of flash memory can be enhanced by hot/cold aware mapping scheme. Given complex meta data and extracting latent user pattern, it is possible to estimate hot/cold data using deep neural network.
Sample Efficient Reinforcement Learning
Currently, Deep Learning architectures suffer in low sample efficiency. Reward propagation is one of the major issues, where the conventional algorithms like DQN update the rewards through single transition. We could notably improve the sample complexity of the agents in various Reinforcement Learning domains.
Compressive Sensing by Deep Learning
Compressive sensing theory provides information-theoretical bound/method to recover signal under the Nyquist sampling rate. Recently, we showed deep learning can outperform conventional compressive sensing algorithms.


In progress