AutoML.org

Maybe it was a long journey, maybe a short one, but most likely not an easy one. Nevertheless, I'm very happy to say that it was definitely a very successful one. Yesterday, Theresa Eimer defended her PhD thesis successfully, and the committee agreed without any discussion that she should be awarded "summa cum laude". Congratulations, Dr. T.! For me, yesterday was also special in another way: Although I co-supervised already three other PhD students and all of them were very successful, Theresa is my first team member in Hannover who graduated from my group. I think we both learned so much over the last 4.5 years from each other. Thank you very much, Theresa, for also showing me the way sometimes and for helping me build up the group in Hannover - especially our RL subgroup. Also many thanks to the other committee members: Joschka Boedecker, Wolfgang Nejdl and Alexander Dockhorn

The world has not ended yesterday and will not today -- but we might have to invest even more energy and time to compensate for certain developments. Therefore I'm very glad to be one of the invited speakers at the (green) AI conference of the Bundesministerium für Umwelt, Naturschutz, nukleare Sicherheit und Verbraucherschutz (BMUV), organized by the Zukunft – Umwelt – Gesellschaft (ZUG) gGmbH. Great discussions about AI for sustainability and more sustainability of AI, with input from researchers, companies and politicians. Later today I will talk about our Green #AutoML project that covers AI development with better models, better hardware and better data encoding combined with AutoML.

ARLBench: Flexible and Efficient Benchmarking for Hyperparameter Optimization in Reinforcement Learning #AutoML and #RL are both considered to be fairly expensive, and combining them is neither trivial nor cheap. Although #AutoRL would be very important to address the problem of brittleness and generalization of #RL, the high computational demand could be one of the reasons why only a few groups work on AutoRL. With ARLBench, we addressed this problem and considerably reduced the compute demand s.t. it is now feasible to do research on AutoRL even if you only have fairly small compute power (or you want to have a smaller CO2 footprint). We achieved this by using efficient JAX implementations, a representative subset of RL benchmarks and a thorough experimental setup. If you want to learn more, check out our paper and GitHub repo (see comments for links). Many thanks to our many co-authors Jannis Becktepe Julian Dierkes Carolin Benjamins Aditya Mohan David Salinas Raghu Rajan Frank Hutter Holger Hoos and Theresa Eimer

Yesterday, I had the pleasure to talk about #AutoML at the conference of the German Service Centers (#KONKIS24). After working on AutoML-related topics for more than a decade, I'm very impressed with how far we are and the great packages that are developed by the AutoML community. I highlighted some of them yesterday -- have a look at my slides. Many thanks also to Marcel Wever (the amazing postdoc in my group -- yes, it's postdoc appreciation week 😁 ) for presenting his work on Sustainable AutoML and to Katharina Strecker for presenting her No-Code AutoML system developed for users in the field of renewable energies.

The #AutoML Conference 2024 is half done, and we now have the second half in front of us 🎉. It is an amazing conference, with great discussions and awesome talks, incl the keynote talks by Chris Van Pelt, Johannes Hoffart, Rich Caruana, Andreas Krause and Mitra Baratchi so far. Each night, there is an event; on Monday, we had a lovely reception at a French Bouillon, and yesterday, we had a joint event with Paris Women in Machine Learning and Data Science, incl. a keynote by Katharina Eggensperger and Alessandra Carbone. Many thanks to Carola Doerr, her team and the Sorbonne Université for having us.

Today, we concluded the 4th #AutoML School. It was my great pleasure to host so many amazing speakers in Hannover, incl. Kristian Kersting, Oleksandr Shchur, Markus Wagner, Setareh Ariafar, Jan van Rijn, Oliver Bringmann, Thomas Meißner, M.A., Katharina Strecker, Frank Hutter, Bernd Bischl, Matthias Feurer, Katharina Eggensperger, Lennart Purucker, Danny Stoll, Neeratyoy Mallik, Christoph Gerum and André Biedenkapp. Also many thanks to my great team for contributing to the success of the event: Marcel Wever, Theresa Eimer, Carolin Benjamins, Tim Ruhkopf, Aditya Mohan, Difan Deng, Leona Hennig and Daphne Theodorakopoulos. With more than 50 attendees, many lectures, keynotes and hands-on sessions, this was a very successful week. In the feedback survey, 92% of the attendees rated the AutoML summer school as "very good" or "good". I'm happy to pass this on and looking forward to the next AutoML school in Tübingen in 2025.

DeepCAVE -- Monitoring and Visualization of #AutoML / #HPO Today, we released the new version of DeepCAVE, our package for monitoring #AutoML and generating a lot of insightful analyses and plots. The new version includes * Support for loading HPO runs generated with AMLTK and Optuna * Capability to import HPO runs from other frameworks by converting them into a Pandas DataFrame * A new plugin for calculating and visualizing ablation path importances * Updated the documentation * Upgraded key dependencies like ConfigSpace, Numpy, and pyPDPPartitioner * Included additional example logs for users to try out the different converters If you like DeepCAVE, please give us a star on GitHub. https://lnkd.in/euYMQxWi Many thanks to my DeepCAVE team, led by Sarah Segel

AutoML Summer School The 4th #AutoML School will take place very soon. It features an amazing lineup of invited speakers talking about the latest ideas in AutoML, but we also have foundational lectures to get you started and even hands-on sessions to learn the practical part. So, we cover the whole range of AutoML. On the website you can also find now abstracts of most invited talks (see comments for link). Important to know is that we only have 3 spots left, and registration closes on Friday (August 23rd) at the latest. So, last call for registration ;-) 

In less than a month, the AutoML Conference 2024 will be in Paris. I don't think that we can quite compete with the #Olympics, but we will give our best ;-) We have an amazing set of tutorials for you. We cover #AutoML tools such as #AutoGluon and #FLAML on Monday and then more research-focus tutorials over the next few days, incl. AutoRL, Meta-Features, Zero-cost proxies for NAS, automatic processing of tables with scikit-learn and skrub. Thanks a lot to the many tutorial speakers preparing these talks, Caner Türkmen, Oleksandr Shchur, Nick Erickson, Li Jiang, Jeff Zheng, Markus Weimer, Theresa Eimer, André Biedenkapp, Gjorgjina Cenikj, Ana Nikolikj, Carola Doerr, Tome Eftimov, Gabriela Kadlecová, Jovita Lukasik and Gael Varoquaux Many thanks to Matthias Feurer, our tutorial chair! PS: The registration is still open, but only for a few more days ;-)

#ICML24 Human-Centered #AutoML Isn't it a contradiction that humans and AutoML should collaboratively work on designing ML applications? AutoML is already taking care of everything? We believe that it is the opposite: Only by combining human expertise and systematic and efficient AutoML, we can leverage the full potential of both. If you are curious about how we envision that concretely, come and visit our poster at [ICML] Int'l Conference on Machine Learning 2024, read our blog post, paper or watch a video about it (Links in the comments).