Hi, I am Guan-Horng Liu (I go by "Guan" ), a third-year Machine Learning PhD in Georgia Tech advised by Evangelos A. Theodorou.
My research aims to develop a new paradigm of Deep Learning Optimization grounded on Optimal Control Theory. This mathematical framework enables rich analysis from stochastic process, game theory, optimal transport, and information duality. It also facilitates principled algorithmic design, better characterization of the training process, and architecture optimization.
I finished my M.S. in Robotics at Carnegie Mellon University, working with George Kantor and Manuela M. Veloso on off-road autonomous navigation and deep reinforcement learning. I also owned a B.S. in MechE at National Taiwan University, with an one-year research exchange at Tokyo Institute of Technology.
I worked in Uber Advanced Technology Group as a Robotics Autonomy Engineer for 1.5 years before joining Georgia Tech, developing motion planning algorithm for self-driving vehicles under the team led by Tony Stentz.
See my full CV here (updated in Feb 2022).
|Feb, 2022||Check out our SB-FBSDE, (accepted to ICLR 2022) on generalizing score-based models with Schrödinger Bridge. Our code is released here!|
|Jan, 2022||I am very fortunate to receive GaTech AE Graduate Fellowship. Thanks GaTech AE!|
|Dec, 2021||I give a talk on Higher-order Optimization of Neural ODEs at DataSig.|
|Nov, 2021||The code for our NeurIPS spotlight is released here!|
|Oct, 2021||I give a talk on Optimal Control Theoretic Neural Optimizer at GaTech ML PhD Seminar.|
|Sep, 2021||I have one paper, SNOpt accepted to NeurIPS 2021 as a Spotlight (acceptance rate 3.0%).|
|May, 2021||I have one paper, DGNOpt, accepted to ICML 2021 as an Oral (acceptance rate 3.0%).|
|Jan, 2021||I have one paper, DDPNOpt, accepted to ICLR 2021 as a Spotlight (acceptance rate 3.8%).|
ICLR 2022Likelihood Training of Schrödinger Bridge using Forward-Backward SDEs TheoryInternational Conference on Learning Representations (* Equal contribution), 2022
NeurIPS 2021Second-order Neural ODE OptimizerAdvances in Neural Information Processing Systems, 2021Spotlight presentation (acceptance rate 3.0%)
ICML 2021Dynamic Game Theoretic Neural OptimizerInternational Conference on Machine Learning, 2021Oral presentation (acceptance rate 3.0%)
ICLR 2021DDPNOpt: Differential Dynamic Programming Neural OptimizerInternational Conference on Learning Representations, 2021Spotlight presentation (acceptance rate 3.8%)
arxivSpatio-Temporal Differential Dynamic Programming for Control of FieldsarXiv preprint (in submission), 2021
RSS 2021Variational Inference MPC using Tsallis DivergenceRobotics: Science and Systems, 2021
arxivA Differential Game Theoretic Neural Optimizer for Training Residual NetworksarXiv preprint, 2020
arxivDeep learning theory review: An optimal control and dynamical systems perspectivearXiv preprint (in submission), 2019
CoRL 2017Learning end-to-end multimodal sensor policies for autonomous navigationConference on Robot Learning, 2017
CMU ThesisHigh Dimensional Planning and Learning for Off-Road DrivingCMU Robotics Institute Master Thesis, 2017
RLDM 2017Multi-modal Deep Reinforcement Learning with a Novel Sensor-based DropoutMulti-disciplinary Conference on Reinforcement Learning and Decision Making, 2017
JBE 2014A bio-inspired hopping kangaroo robot with an active tailJournal of Bionic Engineering, 2014
RSJ 2014Autonomous Control of the WAM-V Catamaran Type Unmanned Surface Vehicle: Propulsion System DesignAnnual Conference of the Robotics Society of Japan, 2014
SII 2013Design of a kangaroo robot with dynamic jogging locomotionProceedings of the 2013 IEEE/SICE International Symposium on System Integration, 2013Best Paper Award