Specifically, I was interviewed on our paper “Explaining Bayesian Optimization by Shapley Values Facilitates Human-AI Collaboration For Exosuit Personalization” (ECML).

The article (in German) discusses the growing market of exosuits and exoskeletons and the importance of personalizing such wearable devices – oftern via Bayesian optimization (BO). This is where our research comes into play: By interpreting BO, taking into account its uncertainty, we allow users to interact with the BO system in a more efficient way. Human-AI collaboration at its best! :rocket:

Here’s a snippet from the article translated to English:

[…] “The statistician Julian Rodemann, now at the Rational Intelligence Lab of the CISPA Helmholtz Center for Information Security in Saarbrücken, worked on Walsh’s project during his doctorate at Ludwig Maximilian University of Munich together with researchers from the Munich Center for Machine Learning. The German research group, in collaboration with the Harvard scientists, developed a mechanism to help users understand why the optimization algorithm recommends certain configurations. “The program offers users of the Back Exosuits two parameters on a scale—one for assistance when bending and one for straightening up—yet it’s very laborious to figure out which combinations are ideal for a given user,” says Rodemann. After all, in theory researchers and users would have to try out every possible pair of values.

“For users to interact with the program more efficiently, they should be able to understand how Bayesian optimization arrives at its suggestions,” Rodemann emphasizes. That’s because it may, for example, also propose values that lie further outside the user’s previous comfort zone precisely because the user hasn’t gathered enough experience there yet. “If the optimization algorithm proposes these parameters without explanation, the user might be more inclined to skip them—and could thereby miss out on a better configuration,” Rodemann explains.

The system is still subject to certain limitations. The idea of interpreting Bayesian optimization directly, rather than the underlying AI model, is still quite new. “We’re able to show to what extent the proposed value for assistance when bending and straightening with the exosuit was suggested for optimization purposes and to what extent it was motivated by a desire to explore the dark corners of the configuration space,” says Rodemann. But of course the user then has to interpret that. “The challenge for programmers now is to translate these values into natural language that’s understandable for every user,” the statistician adds.

Exactly what the user interface will look like remains open. “To find out what kind of information output users respond to best, psychological studies are still needed,” Rodemann underscores. However, the explanation mechanism is slated to be available in upcoming generations of exosuits, which Walsh’s Harvard spin-off is already selling commercially.

Arens concedes that optimizing the walking aids is particularly complex. They must be readjusted not only for different users and tasks but also for different terrains and inclines. The work of engineers, statisticians, and programmers at Harvard and in allied programs such as Munich is an ongoing process: while the spin-off from Walsh’s lab is already selling the first systems commercially, new research is flowing into subsequent generations of the exosuits. The robotics tools are thus intended for increasingly complex practical applications—and to help accelerate the growth of a potential billion-dollar market.

As I just found out, I was nominated as session chair for two sessions (Learning Theory and Optimization). This will be the first time for me to chair a session at a major ML conference, so I’m pretty excited!

This was a huge team effort! Congratulations and a heartfelt Thank you go to Yuanhao Ding, Esteban Garces Arias, Meimingwei Li, Matthias Aßenmacher, Danlu Chen, Gaojuan Fan, Christian Heumann, and Chongsheng Zhang

The award is given to researchers who demonstrate excellence at an early stage of their scientific careers, especially in the field of imprecise probabilities (IP). The prize is generously sponsored by the International Journal of Approximate Reasoning (IJAR).

I received the award (silver :2nd_place_medal:) during last week’s International Symposion on Imprecise Probabilities: Theories and Applications (ISIPTA): https://isipta25.sipta.org/awards

The symposium was a great and inspiring experience. I learned so much! Long poster sessions allowed for in-depth discussion of research in a supportive, open-minded atmosphere :heart: A huge thank you goes to the organizers :pray:

Many, many thanks to the Jury for this recognition! I am extremely grateful to my mentors and collaborators who supported me on this journey: Thomas Augustin, Christoph Jansen, Georg Schollmeyer, Hannah Blocher, Esteban Garces Arias, Malte Nalenz, Dominik Kreiss, Xiao-Li Meng, James Bailie, Yusuf Sale, Eyke Hüllermeier, Matthias Assenmacher, Christian Heumann, Bernd Bischl, Thomas Nagler and many more…

I look forward to continue working on the many facets of imprecise probabilities in statistics and machine learning!

Great collaboration with Federico Croppi, Philipp Arens, Yusuf Sale, Julia Herbinger, Bernd Bischl, Eyke Hüllermeier, Thomas Augustin, Conor J. Walsh, Giuseppe Casalicchio. I will link the paper here once the camera-ready version is available.