Paris-Saclay Center for Data Science

Transcription

1 BALÁZS KÉGL DR / CNRS LAL & LRI CNRS & University Paris-Sud ARNAK DALALYAN CÉCILE GERMAIN ALEXANDRE GRAMFORT AKIN KAZAKÇI Pr / ENSAE Pr / UPSud MdC / Telecom ParisTech MdC / Mines ParisTech Laboratoire de Statistique LRI LTCI CGS 1

2 Meta 2

3 I will not talk about science 3

4 I will talk about management (of) (data) science 4

5 WHERE DOES IT COME FROM? My eight-year of experience interfacing between high-energy physics and data science Our one-year experience of running PSCDS Extensive discussions with management scientist and the MI/Mastodons/MaDICS for the last year 5

6 UNIVERSITÉ PARIS-SACLAY 19 founding partners 6

7 UNIVERSITÉ PARIS-SACLAY + horizontal multi-disciplinary and multi-partner initiatives ( lidex ) to create cohesion 7

8 machine learning information retrieval signal processing data visualization databases Tool building Domain science software engineering clouds/grids high-performance computing optimization human society life brain earth universe A multi-disciplinary initiative to define, structure, and manage the data science ecosystem at the Université Domain scientist Software engineer researchers in 35 laboratories Biology & bioinformatics IBISC/UEvry LRI/UPSud Hepatinov CESP/UPSud-UVSQ-Inserm IGM-I2BC/UPSud MIA/Agro MIAj-MIG/INRA LMAS/Centrale Chemistry EA4041/UPSud Earth sciences LATMOS/UVSQ GEOPS/UPSud IPSL/UVSQ LSCE/UVSQ LMD/Polytechnique Economy LM/ENSAE RITM/UPSud LFA/ENSAE Machine learning LRI/UPSud LTCI/Telecom CMLA/Cachan LS/ENSAE LIX/Polytechnique MIA/Agro CMA/Polytechnique LSS/Supélec CVN/Centrale LMAS/Centrale DTIM/ONERA IBISC/UEvry Neuroscience UNICOG/Inserm U1000/Inserm NeuroSpin/CEA Particle physics astrophysics & cosmology LPP/Polytechnique DMPH/ONERA CosmoStat/CEA IAS/UPSud AIM/CEA LAL/UPSud datascience-paris-saclay.fr Visualization INRIA LIMSI 8 Signal processing LTCI/Telecom CMA/Polytechnique CVN/Centrale LSS/Supélec CMLA/Cachan LIMSI DTIM/ONERA Statistics LMO/UPSud LS/ENSAE LSS/Supélec CMA/Polytechnique LMAS/Centrale MIA/AgroParisTech

9 DATA SCIENCE Design of automated methods to analyze massive and complex data to extract useful information 9

10 CENTER FOR DATA SCIENCE = DATA CENTER We are focusing on inference: data knowledge Interfacing with HPC, cloud, storage, production 10

11 PARAMETERS 2 years: April June 2016, 1.2M +1 year, conditional on evaluation Light management executive committee of 17 members work groups management, strategy (around objectives) thematic (around scientific themes), open to everyone to propose and to participate 11

12 THE DATA SCIENCE LANDSCAPE Data scientist Data engineer Data science statistics machine learning information retrieval signal processing data visualization databases Applied scientist Tool building Domain science Software engineer software engineering clouds/grids high-performance computing optimization energy and physical sciences health and life sciences Earth and environment economy and society brain Domain scientist Data trainer 12

13 CHALLENGES Manpower especially at the interfaces industrial brain-drain Incentives data scientists are not incentivized to work on domain science scientists are not incentivized to work on tools Access no well-developed channels to identify the right experts for a given problem Tools few tools that can help domain scientists and data scientists to collaborate efficiently 13

14 TOOLS We are designing and learning to manage tools to accompany data science projects with different needs 14

15 TOOLS: LANDSCAPE TO ECOSYSTEM Data scientist Data engineer coding sprints Open Software Initiative code consolidator and engineering projects Software engineer Tool building software engineering clouds/grids high-performance computing optimization Data science statistics machine learning information retrieval signal processing data visualization databases Data domains Applied scientist interdisciplinary projects matchmaking tool design and innovation strategy workshops data challenges energy and physical sciences health and life sciences Earth and environment economy and society brain data science bootcamps IT platform for linked data annotation tool SaaS data science platform Domain expert Data trainer 15

16 POSTDOCS, THESES, SABBATICALS Common selection criteria scientific quality expected results both in domain science and data science relevance and feasibility (real) scientific data, available at the start of the project interdisciplinarity (PIs both from domain and data sciences, different LABEXs) community building organizing and participating in thematic days, bootcamps, workshops 16

17 ENGINEERING AND CODE CONSOLIDATING PROJECTS No research implementation, maintenance, integration 1 year engineering projects 3-6 month code consolidating projects a postdoc or PhD student drops research during the project, implements his/her research code in a professional software, or integrates it into a toolbox 17

18 IT PLATFORM FOR LINKED DATA A window to open data at We are not storing or handling existing large data sets Rather indexing, linking, and mapping, embedding in the worldwide linked data (RDF) ecosystem Storing small data sets of small teams is possible Subsets of large sets for prototyping Or simply store metadata plus pointer 18

19 IT PLATFORM FOR LINKED DATA 19

20 BOOTCAMPS Single-day coding sessions participants Goals training PhD students, postdocs, engineers, senior researchers for hands-on data science (problem types, tools) solving (prototyping) real data science problems networking, knowing each other 20

21 BOOTCAMPS 21

22 BOOTCAMPS 22

23 BOOTCAMPS 23

24 BOOTCAMPS 24

25 DATA CHALLENGES A data challenge is a recently developed unconventional dissemination and communication tool a scientific or industrial data producer arrives with a well-defined problem and a corresponding annotated data set defines a quantitative goal makes the problem and part of the data set (the training set) public on a dedicated site data science experts then take the public training data and submit solutions for a test set with hidden annotations submissions are evaluated numerically using the quantitative measure contestants are listed on a leaderboard after a predefined time, typically a couple of months, the final results are revealed and the winners are awarded 25

26 DATA CHALLENGES The HiggsML challenge on Kaggle teams, huge publicity significant improvement on baseline 18 month preparation, yet partially missing the target 26

27 DATA CHALLENGES Challenges are useful for generating visibility in the data science community about novel application domains benchmarking in a fair way state-of-the-art techniques on well-defined problems finding talented data scientists Limitations not necessary adapted to solving complex and open-ended data science problems in realistic environments emphasizes competition 27

28 DESIGN AND INNOVATION STRATEGY WORKSHOPS 28

29 DESIGN AND INNOVATION STRATEGY WORKSHOPS Putting domain scientists, data scientists, and management scientist in the same room Getting them understand each other Keeping them collectively creative The goal: identifying and defining projects low-hanging fruits breakthrough projects long-term vision 29

30 DESIGN AND INNOVATION STRATEGY WORKSHOPS C/K design theory innovative design = interaction and joint expansion of concepts and knowledge 30

31 DESIGN AND INNOVATION STRATEGY WORKSHOPS DKCP process: linearizing C-K dynamics Ini3alisa3on$ [K]$Knowledge$ sharing$ Workshops$ [C]$IFM?Design$ Workshops$ [P]$Project$ building! [RUN]! 31

32 TAKE HOME MESSAGE NO1 If you are interested in adapting any of these tools in your project/site, feel free to contact us, we would be happy to share our experience 32

33 WHAT CNRS CAN DO We need data engineers and trainers to support research ideally: 75% research scientists, 25% research engineers We are lucky in France that the position exists in public research Tasks integrating research code into general-purpose professional software (e.g., scikit-learn, Torch) providing an interface between computational infrastructure (e.g., clouds) and scientists training scientists to use the tools 33

34 WHAT CNRS CAN DO Incentives 34

35 WHAT CNRS CAN DO Most data scientists, as other scientists, are trained and incentivized to do research on highly specialized domains. They search scientific visibility in their international community, which is equally highly specialized, because their carrier advancement is almost entirely based on peer-reviewed publications. Even when they would have the expertise, they have little incentive to venture into the tool builder (data engineer) role since software authorship has little value in their evaluation, and it can only serve them implicitly through the visibility they gain in the community of tool users. By the same token, they have little incentive to venture into domain sciences and to tackle economic or societal challenges. It is possible that a domain science or an industrial project requires new techniques which then can be published in data science venues, but this is not guaranteed at all. It usually takes heavy investment of time and effort to be able to understand domain problems, so excursions into domain sciences are highly risky. Even when such collaborations are established, the data scientist has a strong prior to use his/her favorite methodology which is not necessarily the best solution for a given problem. Finally, the data scientist has little incentive to bring the project to full fruition, and he/she often runs away with an abstract data science problem (and solution) extracted from the project. Symmetrically, domain scientist and industrial domain experts have no incentive to advance data science and to develop and publish new techniques, as long as their data science problems get solved. When they venture into tool development, they have little incentive in developing general purpose tools. 35

36 TAKE HOME MESSAGE NO2 Affirmative action 36

37 TAKE HOME MESSAGE NO2 Affirmative action Overweight (e.g, double count) out-of-domain papers in every evaluation Make software tool ownership count 37

38 TAKE HOME MESSAGE NO3 Co-locality is important. It would be desirable that data science for scientific data in France be grouped into 5-6 sites of similar size. Resources and experience can and should be shared. 38

39 THANK YOU! 39