Section: New Results

Challenges

Participants: Cécile Germain, Isabelle Guyon, Michèle Sebag

PhD: Zhengying Liu, Lisheng Sun

Collaboration: D. Rousseau (LAL), Andre Elisseeff (Google Zurich), Jean-Roch Vilmant (CERN)

Following the highly successful ChaLearn AutoML Challenges (NIPS 2015 – ICML 2016 [106] – PKDD 2018 [45]), the AutoDL challenge [37], to be run in 2019, addresses the problem of tuning the hyperparameters of Deep Neural Networks, including the topology of the network itself. Co-sponsored by Google Zurich, it will require participants to upload their code on the Codalab platform.

In conjunction with AutoDL, we will organize a challenge in computer vision called AutoCV, to promote automatic machine learning for video processing, in collaboration with University of Barcelona. This will make use of the Tau GPU cluster.

Part of the HEP activities of the team, TrackML [30], [61] first phase was run and co-sponsored by Kaggle, until September 2018. The second phase is presently running on Codalab, and will end in March 2019. The challenge has been presented at WCCI [61] and NIPS [30]. I. Guyon and C. Germain are in the organizing committee, and M. Schoenauer is member of the Advisory Committee. The TAU team, in collaboration with CERN, has taken a leading role in stimulating both the ML and HEP communities to address the combinatorial complexity explosion created by the next generation of particle detectors.

Beyond the LAP (Looking At People) series of challenges (see details and references in Section 3.4), the domain of autonomous analysis of human behavior from multimodal information has recently gained momentum. We have been involved in two Special Issues dedicated to these topics, The Computational Face, in PAMI [17], and Apparent Personality Analysis, in IEEE Trans. on Affective Computing [16]. Two other challenges were organized at ICPR 2018, one about the information fusion task in the context of multi-modal image retrieval in social media, the other one regarding the inference of personality traits from written essays, including textual and handwritten information [29].

The HADACA project (EIT Health) aims to run a series of challenges to promote and encourage innovations in data analysis and personalized medicine. The data challenges will gather transdisciplinary instructors (researchers and professors), students, and health professionals (clinicians). The outcome of the data challenges should provide: i) analytical frameworks to bridge the gap between large dataset and personalized medicine in disease treatments and ii) innovative pedagogical methods to sensitize students to big data analysis in health. As a synergistic activity, Tau is also engaged in a collaboration with the Rensselaer Polytechnic Institute (RPI, New-York, USA) to use challenges in the classroom, as part of their health-informatics curriculum.

The L2RPN (Learning to Run a Power Network) project (coll. RTE) [39] addresses the difficult problem of using Reinforcement Learning to assist human operator in their daily tasks of maintaining the French Ultra-High Voltage grid safety while routing power without interruption. We are collaborating with O. Pietquin (Google Brain) to firm up the challenge protocol, largely inspired by AlphaGo and other RL challenges, like the NIPS 2017 “Learning to run” challenge.

It is important to introduce challenges in ML teaching. This has been done (and is on-going) in I. Guyon's Licence and Master courses: some assignments to Master students are to design small challenges, which are then given to Licence students in labs, and both types of students seem to love it. Along similar line, F. Landes proposed a challenge in the context of S. Mallat's course, at Collège de France.