The SAFE IA chair focuses on the theme of safe AI, and more spe­ci­fi­cal­ly on the robust­ness and relia­bi­li­ty of arti­fi­cial intel­li­gence systems.

Presentation

The aca­de­mic in charge of the Chair: Prof. Sébas­tien Destercke

Part­ners: SOPRA STERIA, foun­ding patron of the UTC Foun­da­tion for Inno­va­tion, SCAI (The Sor­bonne Centre for Arti­fi­cial Intel­li­gence), the CNRS and the Uni­ver­si­ty of Tech­no­lo­gy of Com­piègne (UTC).

Insu­ring robust­ness and relia­bi­li­ty is essen­tial in many indus­trial and other appli­ca­tions, such as detec­ting manu­fac­tu­ring defects, obs­tacles in the path of auto­no­mous trans­por­ta­tion, a patient's medi­cal condi­tion, robot control, and more.

The objec­tive of the SAFE AI Chair is to pro­pose new methods to gua­ran­tee the robust­ness of AI models and to test these methods in case stu­dies from the indus­trial world or other appli­ca­tions where relia­bi­li­ty gua­ran­tees are essential.

The chair brings toge­ther mem­bers of five UTC labo­ra­to­ries cur­rent­ly wor­king on AI themes and in fields where AI is due to play an increa­sin­gly impor­tant role in the future. These are the CID team (of which the chair is a mem­ber) at the Heu­dia­syc labo­ra­to­ry and the LMAC labo­ra­to­ry, part of whose work is at the heart of AI, and the UTC’s Rober­val, BMBI, Heu­dia­syc, and Ave­nues labo­ra­to­ries, which cover areas in which AI is set to play a key role. It is also asso­cia­ted with Sopra Steria.

Scientific focus

The chair will com­bine ups­tream research to deve­lop new AI tools capable of respon­ding to exis­ting or future pro­blems encoun­te­red in their appli­ca­tion, with imple­men­ta­tion and inno­va­tion actions based on case stu­dies. To this end, it will main­ly explore three scien­ti­fic areas or issues:

Reliable, safe and trustworthy predictions

The chal­lenge of this research area is to be able to make pre­dic­tions with a gua­ran­teed error rate, with the aim of increa­sing confi­dence in models and moving towards their cer­ti­fi­ca­tion. In par­ti­cu­lar, this area will focus on the pro­blems of pro­vi­ding such gua­ran­tees for each indi­vi­dual, rather than on ave­rage, and of pro­du­cing them in com­plex pre­dic­tion spaces (pro­blems invol­ving a time dimen­sion, images, etc.).

Key­words: cali­bra­tion, sta­tis­ti­cal gua­ran­tees, uncer­tain­ty quan­ti­fi­ca­tion, confor­mable pre­dic­tion, lear­ning with abstention

Appli­ca­tion areas explo­red: Indus­try 4.0 (fault pre­dic­tion), auto­no­mous trans­por­ta­tion (obs­tacle recog­ni­tion), so-cal­led “smart” cities and ener­gy (future consump­tion pre­dic­tion), heal­th­care (medi­cal diagnosis).

Robust models

The chal­lenge of this research area is to obtain models that are robust to imper­fec­tions in the avai­lable data (“small and bad” data rather than “big” data) or to the fact that the deploy­ment envi­ron­ment dif­fers from the model's lear­ning envi­ron­ment, for example when moving from simu­la­tion (in sili­co) or a control­led envi­ron­ment (in vitro) to a real envi­ron­ment (in vivo), or when new classes not present in the lear­ning pro­cess appear.

Key­words: trans­fer lear­ning, robust opti­mi­za­tion, self-lear­ning, opti­mal trans­port, mis­sing or par­tial data, anomaly/novelty detection.

Appli­ca­tion areas explo­red: auto­no­mous vehicle dri­ving and drones (real-world control simu­la­tion), heal­th­care (patient models), Indus­try 4.0 (detec­tion of new faults).

Collaborative learning

The chal­lenge of this area is to improve the qua­li­ty of models, either through model-model col­la­bo­ra­tion (e.g., models based on dif­ferent moda­li­ties or mea­su­re­ments) or through model-human col­la­bo­ra­tion (by soli­ci­ting the expert in a rele­vant and limi­ted manner).

Key­words: co-lear­ning, self-lear­ning, active lear­ning, clas­si­fier fusion.

Appli­ca­tion areas explo­red: e‑health (“smart” homes), “smart” cities and trans­por­ta­tion (mul­tiple sen­sors with pos­sible failure/absence).

Resources

This is a large-scale pro­ject, main­ly fun­ded by the UTC Foun­da­tion for Inno­va­tion and its mem­bers, with human resources pro­vi­ded by all part­ners. At its launch, the chair includes and plans for:

• 5 labo­ra­to­ries, 2 aca­de­mic part­ners, 1 indus­trial part­ner, for a total of more than 15         people (resear­chers, engi­neers, etc.) invol­ved.
• 1 research engi­neer for 3 years
• 6 PhDs
• More than 10 internships
• Invi­ta­tions to forei­gn pro­fes­sors and research scien­tists and/or engineers.

Contacts de la recherche à l'UTC

• UTCéens, ils inventent la com­pote du futur !
• The AI Ren­dez­Vous at UTC : mor­ning lectures
• Mas­ter IDS : La san­té de demain t’attend