Short bio & abstract

Posted 2018.02.21 04:04

세미나 제목

How far can you trust your data and model?
(Incorporating suboptimal data using leveraged Gaussian processes)

Abstract 

When dealing with many machine learning problems, we often assume that the given training data is optimal. However, it may not be the case in general where we only have access to unpolished real-world data. In this seminar, we focus on the problem of incorporating suboptimal or even negative data in a regression problem and its application to robust learning from demonstration. To this end, we present a leveraged Gaussian process (GP) that can model multiple correlated processes in a single GP framework by proposing a leveraged kernel function. The positive definiteness of the kernel is proved using Bochner's theorem. Given a leveraged GP framework, we further present a sparse-constrained leverage optimization method that can optimize the leverage (proficiency level) of each training data inspired by the model selection problem. The sparse constraint effectively overcome the inherent ill-posedness of finding correlations of each training data between Gaussian processes. The proposed methods have been successfully applied to a robust learning from demonstration problem where demonstrations with mixed qualities are given without labeling. We further present our recent work of estimating the uncertainty of a prediction when using a deep neural network. Particularly, we present a sampling-free uncertainty modeling method using a mixture density network by decomposing the uncertainty into explained and unexplained variances and show our preliminary result on a real-world driving dataset. 

Short Bio

Sungjoon Choi is a Ph.D. candidate at the Seoul National University in the Department of Electrical and Computer Engineering where he received the B.S. degree in Electrical Engineering from the Seoul National University in 2012. He is currently a research intern at Disney Research Los Angeles. His current research interests include stochastic processes, modeling predictive uncertainty in deep learning, and nonparametric machine learning algorithms with applications to robotics. Choi received Best Conference Paper Finalist Award at 2016 IEEE International Conference on Robotics and Automation. 







Sorry about the late reply, and yes! I am very excited as well of going back to DRLA for the job talk. 
Requested information are as follows:

1) Talk Title: Robust learning from demonstration and its application to robotics

2) Abstract: 

In this talk, we will briefly review Gaussian process regression and introduce our leveraged Gaussian process that can handle multiple correlated processes in a single GP framework. Leveraged Gaussian processes can handle sub-optimal (or suscpicious) training examples instead of excluding them from the training set. We will cover robust learning from demonstration where we may have a mixture of positive and negative demonstrations and its application to navigation and human-robot-cooperation tasks. In addition, we will cover how to effectively apply reinforcement learning using real hardwares and estimate the uncertainty information in predictions from the complex machine learning model.

3) Short bio:

Sungjoon Choi received Ph.D. in Electrical and Computer Engineering from Seoul National University (2018), MA in Electrical Engineering from Seoul National University (2012). He is currently a research scientist at Kakao Brain in Korea and was a research intern at Disney Research Los Angeles in 2017. His current research interests include reinforcement learning, imitation learning, human-robot-interaction, and modeling uncertainty in deep learning. Choi received Best Conference Paper Finalist Award at 2016 IEEE International Conference on Robotics and Automation (ICRA) and published more than 10 papers to conferences and journals in robotics as a first author. 

4) Yes! It is my latest CV. 

I have quick questions about the interview. Will it be possible to know about the detailed timeline of the interview? For example, when should I arrive at the campus and how long will the interview take? 

Thanks again for reminding me. 

Best,


'Thoughts > Technical Writing' 카테고리의 다른 글

ICLR 2018 reviews  (0) 2018.11.22
ICML 2018 Review  (0) 2018.04.11
Short bio & abstract  (0) 2018.02.21
Postdoc 자리를 물어보는 이메일  (0) 2017.12.16
영어 논문 글쓰기  (0) 2017.07.03
Reviews I got from IROS 2017  (0) 2017.06.27
« PREV : 1 : ··· : 30 : 31 : 32 : 33 : 34 : 35 : 36 : 37 : 38 : ··· : 635 : NEXT »