Uni-Logo

Frank Hutter

Frank Hutter

 Professor

Postal address:
Institut für Informatik
Albert-Ludwigs-Universität Freiburg
Sekretariat Hutter/Maschinelles Lernen
Georges-Köhler-Allee 074
79110 Freiburg, Germany
Room:
Building 074, Room 00-017
Coordinates:
48.014472, 7.831111 (DD)
Email:
fh@cs.uni-freiburg.de
Phone:
+49 761 203-67740
Fax:
+49 761 203-74217

I'm the Head of the Machine Learning Lab. I'm lucky enough to have an amazing team; all my team members are linked in the sidebar on the left. In addition to my full-time role at the University of Freiburg, I also consult for the Bosch Center for AI (BCAI) as Chief Expert for AutoML.

Information for students interested in projects, theses, or Hiwi positions

With machine learning being one of the hottest topic around, our small group is flooded with requests. To make the process efficient, please do not email me directly, but follow the instructions posted here.
Recent News
Major research project on AutoML funded by Bosch. The Bosch Center for Artificial Intelligence (BCAI) is funding my fundamental research on automated machine learning (AutoML) with 6.4M Euro. Here is the university's press release. Independently of this fundamental research, I will also consult for Bosch to help them bring AutoML to bear in practice.
I taught tutorials at both ICML and CVPR. ICML tutorial on algorithm configuration, with Kevin Leyton-Brown: slides, video. CVPR tutorial on meta-learning, with Nikhil Naik, Chelsea Finn, and Nitish Keskar: slides. We also organized another great edition of the AutoML workshop.
3/3 ICLR papers accepted! 1. Learning to Design RNA, 2. Decoupled Weight Decay Regularization, and 3. Efficient Multi-Objective Neural Architecture Search via Lamarckian Evolution. The last of these is on neural architecture search (NAS), and our NAS survey has also been accepted for publication in JMLR.

New book on AutoML! Along with Lars Kotthoff and Joaquin Vanschoren, I edited a new book on AutoML, published with Springer. The book is fully open access, but hard copies can be ordered starting February 2019. The book contains reviews on hyperparameter optimization, neural architecture search, and meta-learning, descriptions of prominent AutoML systems and a review of AutoML challenges.
I gave a tutorial on AutoML at NeurIPS 2018. A recording is available here, and slides are available here, with all references being clickable links. My co-speaker was Joaquin Vanschoren. I also coorganized the MetaLearning workshop again at NeurIPS.
We did it again! World champions in automatic machine learning (AutoML)! After winning the first international AutoML challenge 2015-2016, we also just won the second international AutoML challenge, which ran 2017-2018. This second AutoML challenge was again organized by ChaLearn and was affiliated with PAKDD 2018. Our system POSH-Auto-sklearn outperformed all other 41 AutoML participating systems. (Back in 2016, our AutoML system also won against human experts, but this time there was no human track.)

I received a Google Faculty Research Award. I am one of only 14 researchers worldwide selected for a Google Faculty Research Award in the area of Machine Learning and Data Mining in 2017-2018, and the only one in Europe. I will use the award to develop new state-of-the-art methods for efficient joint neural architecture search and hyperparameter optimization, building on our existing work on efficient Bayesian optimization, learning curve extrapolation, and Auto-Net.

General research interests

I am interested in all facets of intelligence, and how we can replicate it in artificial systems. In particular, I work on
  • Statistical machine learning (in particular deep learning), to learn effective representations for large amounts of very noisy data (including uncertainty quantification)
  • Automated problem solving, including knowledge representation
  • Autonomously-learning software systems, which can improve their performance over time without the need for a human in the loop
  • Sequential decision making under uncertainty, to trade off building better models of the world vs. acting better based on them
  • Scientific experimentation, to make empirical research more reproducible and to codify human experts' strategies to the point where an autonomous system can execute them.
Most of my research is of a methodological nature on the meta-level, such as machine learning methods that model & optimize the performance of other machine learning methods. These meta-level problems constitute a great `drosophila' for AI: they require representation learning, reasoning under uncertainty, bounded rationality, active learning, exploration/exploitation tradeoffs, transfer learning, and much more, but they also allow for a perfectly reproducible environment with complex dynamics and available ground truth (given enough computational power). Also, they are great from a practical perspective since their outputs are better algorithms for problems we care about.

Specific research areas

I currently focus on a few research areas that combine the general themes above:
  • Bayesian optimization -- sequential experimental design under uncertainty
  • Automated machine learning -- developing an AI that can compete with human data scientists
  • Deep learning -- automatically learning representations of the data; in particular, I work on automating structure & hyperparameter search for deep learning, and on improving optimization algorithms for deep networks
  • Automated algorithm design -- developing automated methods for parameter optimization, algorithm selection, and algorithm analysis. A few years ago, I gave a Google tech talk on this topic; you can watch it on youtube.

Affiliations

I'm a member of the Department of Computer Science at the Faculty of Engineering of the University of Freiburg. In addition to my full-time role at the University of Freiburg, I also consult for the Bosch Center for AI (BCAI) as their Chief Expert for AutoML. I hold an ERC Starting Grant from the European Research Council and an Emmy Noether Grant from the German Research Foundation (DFG). I'm a former member of the Computer Science Department of the University of British Columbia (UBC), specifically of the Laboratory for Computational Intelligence (LCI) and the Bioinformatics and Empirical & Theoretical Algorithmics Laboratory (BETA). I was a machine learning consultant for Zynga Inc and am a co-founder of Meta-Algorithmic Technologies. I earned my PhD at UBC in 2009 and my Diplom (eq. MSc) at Darmstadt University in 2004.

I will be program co-chair of ECML 2020 in Ghent. I co-founded and regularly co-organize the NeurIPS workshop series on meta-learning and the ICML workshop series on AutoML. I also co-founded and regularly co-organized the NeurIPS workshop series on Bayesian optimization. Please see http://automl.org/workshops for an up-to-date list of workshops I'm co-organizing..

I'm currently on the editorial board of JAIR and am regularly area chair, senior programme committee member or reviewer for NeurIPS, ICML, and ICLR. I have also served in these roles for IJCAI and AAAI.

Please see my publication page and our website http://automl.org for our blog, research topics, events we organize, etc. My academic CV (the CV is updated very irregularly).

Misc

Erdös number: 3 (Anne Condon -> Michael E. Sacks -> Paul Erdös)

Tweets


For finer-grained news, here are some of my tweets: