Rein Houthooft

Research Scientist at OpenAI
in artificial intelligence / machine learning

About

After graduating summa cum laude from Universiteit Gent, Belgium in 2014 with an engineering degree (burgerlijk ingenieur) in computer science, I obtained a doctorate in the beginning of 2017 at IDLab, which was supported by a Doctoral Fellowship of the Research Foundation — Flanders (FWO) for fundamental research.

Initially I performed applied research in combinatorial optimization and operations research at the CODeS group, part of KU Leuven, Belgium, in collaboration with ArcelorMittal. Hereafter I worked in the domain of network science, designing experimental routing strategies for computer networks. Later my interests shifted to artificial intelligence and machine learning, which became my true passion.

More specifically I focused on combining deep neural networks with structured prediction. During my doctoral studies, I also acted as a researcher at the R&D institute imec, which allowed me to apply my research to visual perception in autonomous agricultural vehicles, in collaboration with CNH Industrial.

Later as part of my doctoral studies, I investigated exploration strategies for deep reinforcement learning. During this time, I pursued research both at OpenAI and at the Berkeley Artificial Intelligence Research lab (BAIR), part of UC Berkeley, with Pieter Abbeel. This latter stay was supported by a Travel Grant of the FWO.

Since 2017, I am a research scientist at OpenAI in machine learning with a focus on deep reinforcement learning. I am co-organizing the Deep Reinforcement Learning Symposium at NIPS and was involved in the Berkeley Deep RL Bootcamp.


Projects



Deep Reinforcement Learning and Generative Models

Reinforcement learning (RL) using nonlinear function approximators with a focus on continuous control tasks such as robot locomotion. In particular, the goal is to investigate how to achieve efficient exploration in deep RL through curiosity. This research was performed in collaboration with OpenAI and the Berkeley AI Research lab.

Parameter Space Noise for Exploration
arXiv, 2017
M. Plappert, R. Houthooft, P. Dhariwal, S. Sidor, R. Y. Chen, X. Chen, T. Asfour, P. Abbeel, M. Andrychowicz
#Exploration: A Study of Count-Based Exploration for Deep Reinforcement Learning
Advances in Neural Information Processing Systems (NIPS), 2017
H. Tang, R. Houthooft, D. Foote, A. Stooke, X. Chen, Y. Duan, J. Schulman, F. De Turck, P. Abbeel
VIME: Variational Information Maximizing Exploration
Advances in Neural Information Processing Systems (NIPS), 2016          
R. Houthooft, X. Chen, Y. Duan, J. Schulman, F. De Turck, P. Abbeel
InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets
Advances in Neural Information Processing Systems (NIPS), 2016
X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, P. Abbeel
Benchmarking Deep Reinforcement Learning for Continuous Control
International Conference on Machine Learning (ICML), 2016
Y. Duan, X. Chen, R. Houthooft, J. Schulman, P. Abbeel




Structured Prediction and Deep Learning

As part of an autonomous vehicle project, the goal was to combine structured output prediction and deep learning techniques, with a particular focus on semantic image segmentation. Structural support vector machines (SSVMs) were extended to allow for highly nonlinear factors. This can enhance output coherence of deep predictive models, while still allowing for end-to-end training. Below the architecture of a deep SSVM with convolutional neural factors is pictured.


Integrated Inference and Learning of Neural Factors in Structural Support Vector Machines*
Pattern Recognition, vol. 59, 2016
R. Houthooft, F. De Turck
Structured Output Prediction for Semantic Perception in Autonomous Vehicles*
The 30th AAAI Conference on Artificial Intelligence (AAAI), 2016
R. Houthooft, C. De Boom, S. Verstichel, F. Ongenae, F. De Turck

* This work is part of an applied research project in collaboration with Case New Holland (CNH) Industrial. As such several methods, models, datasets, and results could not be publicly released due to confidentiality agreements. An addendum to these papers can be found here. Initial vehicle controller patent applications have been filed.





Artificial Intelligence in Healthcare

In collaboration with the Laboratory of Intensive Care Outcomes Research of the Universitair Ziekenhuis Gent, this applied research project investigates the use of machine learning techniques for predictive modeling and medical decision support in the intensive care unit.

Predictive Modelling of Survival and Length of Stay in Critically Ill Patients using Sequential Organ Failure Scores
Artificial Intelligence in Medicine, vol. 63, no. 3, 2015
R. Houthooft, J. Ruyssinck, J. van der Herten, S. Stijven, I. Couckuyt, B. Gadeyne, F. Ongenae, K. Colpaert, J. Decruyenaere, T. Dhaene, F. De Turck




Network Science

My research originally focused on the development of a novel routing algorithm called Forest Routing. Through geometric routing, using a set of graph embeddings in a particular mathematical space, it offers both high scalability and native load balancing behavior. A coherent write-up on the subject can be found in my thesis Adaptive Geometric Routing for the Internet Backbone. Below a demonstration of the developed model is shown.



Optimizing Robustness in Geometric Routing via Embedding Redundancy and Regeneration
Networks, vol. 66, no. 4, 2015
R. Houthooft, S. Sahhaf, W. Tavernier, F. De Turck, D. Colle, M. Pickavet
Robust Geometric Forest Routing with Tunable Load Balancing
The 34th Annual IEEE International Conference on Computer Communications (INFOCOM), 2015
R. Houthooft, S. Sahhaf, W. Tavernier, F. De Turck, D. Colle, M. Pickavet
Fault-Tolerant Greedy Forest Routing for Complex Networks
The 6th International Workshop on Reliable Networks Design and Modeling (RNDM), 2014 — best paper award
Featured in Global Communications Newsletter May 2015
R. Houthooft, S. Sahhaf, W. Tavernier, F. De Turck, D. Colle, M. Pickavet



Contact

Feel free to contact me through any of the below.

rein.houthooft@gmail.com