3iA Activity Report 19-21 by Université Côte d'Azur

AI for health and smart territories Activity Report 2019-2021

Table of contents INTRODUCTION •

Editorial p4

•

What’s 3IA Côte d’Azur ? p.5

•

The 3IA national network p.6

•

Local dynamic p.7

•

3IA Côte d'Azur in numbers

p.8

RESEARCH P.10 •

Scientific publications p.13

•

Core element of AI p.17

•

AI for integrative computational medicine

p.25

•

AI for computational biology and bio-inspired AI

p.29

•

AI for smart and secure territories

p.33

EDUCATION AND TRAINING P.38 •

AI for specialists

•

AI for non-specialist students

•

AI for continuing education in companies

•

AI for citizen outreach

PARTNERSHIPS AND INNOVATION p.44 •

Socio-economic impact

•

Valorization and development

GOVERNANCE AND BUDGET p.56 INTERNATIONAL VISIBILITY p.51 WHAT’S NEXT ? p.62 APPENDIX: p.65 •

Research unit partners of the project

•

List of 3IA PhD's and Post-docs

Editorial by the Director The Institut 3IA Côte d'Azur, launched in September 2019 and coordinated by Université Côte d'Azur, is dedicated to the applications of Artificial intelligence (AI) to health and for smart territories. After two years of existence, a great dynamic is taking place around AI on Côte d'Azur, and the Institute is positioning itself as a tool for its ecosystem in which it plays a driving role, bringing together business, research and training institutes and public players. The 3IA Côte d'Azur is one of the 4 Interdisciplinary Institutes of Artificial Intelligence labeled in France in April 2019. The institute is carried by six key academic actors of AI in the region: Université Côte d’Azur, CNRS, Inria, Inserm, Eurecom and Skema. The project is also supported by Mines ParisTech, Nice University Hospital, INRAE, local authorities and more than 100 companies and startups. The institute has a total budget of more than 50 million euros for the period 2019-2023. These first two years of exercise of the Institute have already been crowned with great achievements in structuring the AI ecosystem, in research (more than 500 scientific publications), in innovation (more than 10 million Euros invested by local companies in research collaborations) and education (+50% of people trained in AI, with the launch of new initial and continuing training programs). This first report allows both to make a progress review on the deployment of this new model of interaction between academic research, companies and society, and to project the action of the institute in a longer term. More importantly, the Institute has established itself as a federating player and a local entry point for transforming AI technological innovation emerging from research laboratories into use cases for companies and local authorities, while supporting all citizens to train or acculturate to these new technologies. The Institut 3IA Côte d'Azur is also deeply involved, in close collaboration with the other 3IA institutes, in the French AI strategy "AI for Humanity", for which a new phase1 with 1.5 B€ has been recently announced, and the development of the image of the French research in AI at the international level. On the strength of this excellent collective dynamic on the Côte d'Azur and at the national level, the Institute is now in full capacity to carry out its research, innovation and training missions, and can project itself actively for the years to come. Pr. Charles Bouveyron Director of 3IA Côte d'Azur

https://www.gouvernement.fr/investissements-d-avenir-nouvelle-phase-de-la-strategie-nationale-d-intelligence-artificielle-le

What’s 3IA Côte d’Azur ? / A CONSORTIUM The Institute started on September 1, 2019 and is coordinated by Université Côte d’Azur. It is a consortium composed of Université Côte d’Azur (“UCA”- coordinator), the National Institute for Research in Digital Science and Technology (Inria), the French National Center for Scientific Research (CNRS), EURECOM, the French National Institute of Health and Medical Research (INSERM), and SKEMA Business School. The project is also supported by MINES ParisTech, the Nice University Hospital, INRAE, local authorities and, at project start, by 62 companies.

/ HEALTH AND SMART TERRITORIES

/ BUDGET OF THE INSTITUTE

Four research axes : The overall budget of the Institute was €49.5M over 4 years when the 1. Core element of AI project was approved. With SKEMA 2. AI for integrative computational medicine 3. AI for computational biology and bio-inspired AI joining the consortium, the global budget now amounts to €51.5M 4. AI for smart and secure territories

/ MISSIONS •

Make AI expertise, algorithms and tools avail-

•

Create an innovative ecosystem that is in-

able to academia, companies, and the com-

fluential at the local, national and inter-

munity, in order to have a significant impact in

national levels, and a focal point of ex-

terms of economic value but also in the every-

cellence

day life of the citizens.

for

research

and

education.

/ THE 3IA NATIONAL NETWORK At the national level, the network of 3IA Institutes

second edition of the “AI4Health’’ winter school

plays a central role in the French national AI policy

was held (online) on January 12-14, 2022.

“AI for Humanity’’, launched in the Fall 2018 by President Macron. After a first year during which

Now, with the support of INRIA, which coordinates

the

Institutes

the national research program, the network of 3IA

built their own

Institutes has reached cruising speed with regular

local models, the

meetings of the management teams, common

network of 3IA

working groups (education, medical imaging, etc.),

Institutes was really activated in 2020 with the first

an effective website to work and communicate

joint actions.

and regular joint events, such as the colloquium for PhD students of the four 3IA Institutes (first

The network of 3IA Institutes organized several

meeting held in November 2021, hosted by ANITI

events early in 2020, such as a “Diversity” day and

in Toulouse).

a round table “Towards an egalitarian AI?”, which took place on March 8, 2020, a working group on

The network of 3IA Institutes is also currently work-

AI teaching in the Spring 2020 and the first edition

ing, at the request of Mme Fredérique Vidal, French

of the “AI4Health’’ winter school which took place

Minister of Higher Education and Research, on de-

in January 2021.

signing the roadmap for the Education program

The latter action is the result of a strong coordina-

(€800M over 5 years) of the second phase of the

tion between the 3IA Institutes MIAI, Prairie and 3IA

French national plan for AI. The objective here is to

Côte d’Azur on the topic of Health, in collaboration

standardize the model built by the 3IA Institutes at

with the French “Health Data Hub’’ initiative. The

the national level.

www.instituts-3IA.fr

/ LOCAL DYNAMIC AND COMMUNICATION TO SOCIETY Based within Sophia Antipolis, Europe’s leading technology park, the Institute’s actions are coordinated with those of its partners (academics, local authorities, companies, competitiveness clusters, technology transfer subsidiaries) and it plays a driving and central role within its ecosystem, strengthening its dynamism and visibility. The creation of 3IA Côte d’Azur has also led to the creation of new actors linked to AI and digital technology in the region, which the Institute works with very closely. A first example is the “Maison de l’Intelligence Artificielle”, dedicated to the education of school children, founded by the Alpes-Maritimes Department, the Sophia Antipolis Urban Community (CASA), the Nice Côte d’Azur Chamber of Commerce and Industry and Université Côte d’Azur. An observatory of the economic and societal impacts of AI, named OTESIA, has also been created in the wake of the Institute. OTESIA conducts studies to measure the impact of AI in different fields and to prepare specific fields, such as education, to the use of AI. Finally, the EuropIA association, also created in 2019 at the initiative of the Alpes-Maritimes Department, aims to develop events around AI for the general public. The creation of the Institute has

enabled

the

local

companies working on AI to form networks so as to interact effectively with the Institute on their AI needs and the orientation of their R&D strategies. The “Industrial Research Council for Artificial Intelligence” (ICAIR) group was launched in June 2019 and gathers large French and international groups, such as Accenture, ACRI-ST, AIR FRANCE-KLM, Amadeus, ARM, Hewlett Packard Enterprise, IBM, NXP Semiconductors, Orange, Renault Software Labs, SAP Labs France, STMicroelectronics and Thales Alenia Space. The “Cluster IA”, founded in the Spring 2019 by Docaposte and the startup Synchronext, along with Université Côte d’Azur, brings together the SMEs & startups of the AI ecosystem of the Côte d’Azur region. 3IA Côte d’Azur is therefore now establishing itself as a central element in its region and a natural point of contact for all requests relating to AI (R&D partnership, creation of tailor-made training courses to meet the needs of companies, co-organization and participation in events related to AI, etc.). The Institute also contributes to all these associations and groups by the presence of chairholders, PhD students or post-doctoral researchers at various events and actions aimed at companies and young members of the general public.

3IA Côte d’Azur in Numbers (2019-2021) GOVERNANCE AND BUDGET • €51.5M : overall budget of the Institute • €19.5M from the public partner institutions • 6 founding members

EDUCATION AND TRAINING A large number of training programs including

20 certifications with an AI content >50% 44 PhD students to teach 32 hours of AI per semester • +50% increase in the number of people trained

•

From the beginning of 2019 until 2021, the teaching potential of our PhD students increased from 640 hours (1st academic year), to 1,408 hours (2nd academic year), to 2,560 hours (current academic year).

•

AI FOR SPECIALISTS +1,100 people / year AI FOR NON-SPECIALIST STUDENTS +1,200 people / year AI FOR CONTINUING EDUCATION IN COMPANIES +300 people/year AI FOR CITIZEN OUTREACH: +3,000 people/year

• €160k of profits generated by professional training in 2021

• •

The Institute human resources: +65 people

in 3 years

• •

RESEARCH • 47 Chairs in 4 strategic scientific areas including 3 International Chairs including 4 affiliate chairs • 44 PhD students and 16 post-doctoral researchers PUBLICATIONS • 484 scientific publications since 2019 • + 20% per year compared to the 2014-2019 period

•

PARTNERSHIPS AND INNOVATION 140 companies which the Institute is in contact with 57 contracts signed €4.27M of cash income €10.8M of financial investment from companies

Patents • •

13 software patents 8 industrial patents

Start-up • •

3 running 3 more scheduled

3IA TechPool •

3 engineers to perform software development tasks

Study Group for Industry • • •

5 companies 5 problems to solve 25 students involved

6 Webinars for companies • •

150 participants 70% of companies

RESEARCH

AI for real world, with applications in healthcare and smart territories

The research activity of the Institute is based on the idea that AI operates in the real world, particularly with applications in healthcare (medicine and biology) and smart territories. The scientific program covers four main research areas: core elements of AI, AI for integrative computational medicine, AI for computational biology and bio-inspired AI, and AI for smart and secure territories. The activity the each chairholder gravitates around these four axes and is not limited to the single axis the chair is related to. Synergies between the chairholders and their teams across the different axes show that cross-fertilization starts from the core elements of AI and leads to the more applicative axes in computational medicine (e.g., the federated learning FedBioMed project of M. Lorenzi and the ANTIDOTE project of E. Cabrio and S. Villata), in biology (e.g., the single cell and molecular studies on SARS-CoV-2 of P. Barbry and F. Cazals), and in smart territories (e.g., the traffic flow improvement of P. Goatin and the smart building of P. Alliez). These more applicative axes generate new challenging issues to be addressed as fundamental AI problems in Axis 1 (e.g., cyber-defense through statistical network analysis by C. Bouveyron). We provide a summary below of the implementation of the research program and its results.

A successful research strategy A pool of high level reseachers The Institute currently comprises 47 Chairs in 4 strategic scientific areas: •

Core elements of AI,

•

AI for integrative computational medicine,

•

AI for computational biology and bio-inspired AI,

•

AI for smart and secure territories.

The Institute started in 2019 based on the 27 Chairs selected ab initio by the international jury, as well as 3 international chairs. In order to strengthen Axis 3 “Computational

Evolution of the number of academic Chairs by axis

biology

and Bio-inspired AI’’ and Axis 4 “Smart Territories”, 13

new Chairs have been awarded to researchers from the Institute community. Finally, 4 affiliated chairs were also awarded in 2020 and 2021 to researchers from partner companies. These researchers devote 20% of their time to the Institute’s research, benefit from 3IA’s scientific activities and may

Number of academic Chairs

40 academic chairs 3 international chairs 4 affiliated chairs

16 14

12 10

8 6

2 7

4 2 0

Axe 1

Core element of AI

Axe 2

Axe 3

Int. Comput. Medecine

2019

Int. Comput. Biology

2020

Axe 4

Smart and Sec.Terr.

2021

co-supervise PhD students or post-doctoral researchers. These affiliated chairs strengthen ties with partner companies. With these new recruitments, a total of 47 chairs are involved in 3IA research as of December 2021. As expected, the work of the chairholders has been strongly supported by the Institute through the recruitment of PhD students and post-doctoral researchers. The Scientific Council of 3IA organizes three calls for applications per year for these recruitments. Thus, 44 PhD students and 16 post-doctoral researchers have been recruited. Notice that 6 of these 44 PhD students have been recruited to work with researchers who are not 3IA chairholders. This is a strong indication that the Institute action is not limited to Chairs only. In addition, the Institute is developing a program of visiting researchers to

44 PhD's 16 Post-docs

strengthen its attractiveness and its international links. Funding for 4-month invitations had been earmarked for foreign researchers but only a part (3 over 7) has been honored because of the pandemic, but the system will be relaunched as soon as the situation allows it.

481 Scientific publications since 2019 A page of the French Open Science website is dedicated to our Institute: https://hal.archivesouvertes.fr/3IA-COTEDAZUR/ It allows users to browse and access all publications since the beginning of the Institute. First, regarding publications, the dedicated page of the French open science website HAL (hal.archives-ouvertes.fr/3IACOTEDAZUR/), which lists all the publications of the Institute from

Progression in the number of publications by researchers who currently hold a 3IA Côte d'Azur Chair, for the period 2014-2021

its start, shows the number of publications by chairholders in-

300

hhh the top-tiered conferences (Neurips, ICML, AISTATS, IJCAI,

250

ECAI, AAAI, CVPR, ICCV, and MICCAI) and in artificial intelligence

200

journals (Nature journals, Machine Learning, AI in Medicine,

150

NeuroImage, IEEE PAMI, Annals of Applied Statistics, etc.).

100

To date, the activity of the Institute’s researchers has produced

484 scientific publications since 2019. These include a large

number of interdisciplinary publications, some co-authored

2014 Axis 1

2015

2016

2017

Axis 2

2018

Axis 3

2019

2020

2021

Axis 4

Figure 1

with companies (Criteo AI Lab, Frog Labs AI San Francisco, Amadeus, SAS Institute Inc., Toyota Motor Europe (Belgium), and Dassault Systèmes).

The number of publications in AI by researchers who currently hold a chair has significantly increased since the Institute was recognized in 2019, compared to the previous five years. The Institute community generated 20% more publications

200

per year compared to the 2014-2019 period. Second, it is

+ 20% per year compared to the 2014-2019 period

important to note that the Institute improved interdiscipli-

150

100

narity, in particular between health applications of AI and

smart territories. The FedBioMed project and the several projects on the fight against Covid-19 are representative of this integration. Finally, the Institute has already had a great impact on research collaborations with private companies.

0 Journals

Conferences

Books

Others

Figure 1 : Share of the publications of 3IA Côte d’Azur researchers by types and years Figure 2

Interestingly, the Axis 1 “Core elements of AI” and Axis 4 “Smart territories” generate, often jointly, a large part (84%) of all research contracts.

Figure 1 highlights the positive trend established by the Institute since its creation, with an increasing number of publications by the chairholders and their teams under the 4 research axes of the Institute. Figure 2 shows the number of publications the Institute has accumulated since its creation.

Publications in top-tier conferences and journals in AI Results from the chairholders and their teams have been published in top-tier conferences and journals in artificial intelligence, including IJCAI (3), ECAI (2), AAAI (1), CVPR (4), ICCV (3), MICCAI (7), ISWC (1), AAMAS (2), COLING (1), Artificial Intelligence in Medicine (1), NeuroImage (6), IEEE Transactions on Signal Processing (4), Medical Image Analysis (8), IEEE Transactions on Medical Imaging (1), and IEEE Transactions on Pattern Analysis and Machine Intelligence (1), as well as machine learning, for example, ICML (4), AISTATS (2), NeurIPS (7), and Annals of Applied Statistics (1). Three articles were published in Nature journals : a survey on AI in cardiovascular imaging in Nature Reviews Cardiology involving the 3 Chairs of Sermesant, Delingette, and Ayache, one publication in Nature Medicine, and one in Nature Communications. Also, Ayache published a special issue and editorial in the Proc. of the IEEE. He was the French coordinator for the Science Academies of the Group of Seven (G7) Nations who wrote a statement in 2021 on "Data for international health emergencies: governance, operations and skills".

The chairholders also co-authored 4 books, namely: •

C. Bouveyron, G. Celeux, B. Murphy, and A. Raftery, Model-based Clustering and Classification for Data Science, with Applications in R, in Series in Statistical and Probabilistic Mathematics, Cambridge University Press, 2019.

•

D. Allemang, J. Hendler, and F. Gandon. Semantic Web for the Working Ontologist. ACM, 2020.

•

A. Betti and M. Gori, Deep Learning to See: Towards New Foundations of Computer Vision, Springer, 2021.

•

X. Pennec, S. Sommer, and T. Fletcher. Riemannian Geometric Statistics in Medical Image Analysis. Elsevier, 2020.

Awards and recognition The chairholders of the Institute and their teams have received several national and international awards, and recognition certifying the excellence of their research achievements. In particular, we highlight the following: M. Filippone (AXA Chair of Computational Statistics 2016-2023), N. Ayache (International Steven Hoogendijk Award 2020), D. Wales (Humboldt research prize of the Alexander von Humboldt foundation in 2020), F. Delarue (AMS Doob prize in 2020 and SMF/SMAI Maurice Audin prize in 2020), P. Reynaud-Bouret (CNRS Silver Medal in mathematics in 2021 and Pierre Faure prize of the French Academy of Sciences in 2020), R. Deriche (EURASIP Fellow 2019), S. Ourselin (Fellow of the Royal Academy of Engineering FREng, Fellow of the Institute of Physics and Engineering in Medicine FIPEM), D. Gesbert (Research grand prize of IMT–French Academy of Sciences [MinesTélécom] 2021), and S. Villata (Junior Researcher Award Inria of the French Academy of Sciences 2021).

Invited talks The high-quality research addressed by the chairholders and their teams also resulted in several invitations to give invited talks and lectures at prominent national and international venues. Among others, we may highlight the invited talks of N. Ayache (Global Forum on AI for Humanity 2019, French German AI Symposium 2020, Int. PRAIRIE workshop 2021), C. Bouveyron (ECDA 2020, FGML 2021, IFCS 2022), F. Delarue (International Congress of Mathematicians 2022), X. Pennec (Computational Geometry Week 2020, Recent Advances in Statistical Analysis of Imaging Data of the American Statistical Association 2020), S. Villata (INTAP 2020, ENDORSE 2021), D.Wales (Theoretical Chemistry Symposium 2021), L. Blanc-Féraud (IEEE CAMA 2021), and F. Bremond (IPAS 2020).

Organization of international scientific events Several international scientific events have been organized by the chairholders. Among others, an AMS Short Course in Denver on “Mean Field Games: From Agent Based Models to Nash Equilibria” in 2020 (F. Delarue), two international workshops at NeurIPS 2019 and NeurIPS 2021 on “Optimal Transport for Machine Learning” (R. Flamary), the NeurIPS Workshop on “Beyond BackPropagation: Novel Ideas for Training Neural Architectures” (M. Gori), the special session on “Security and Fairness in Collaborative Healthcare Data Analysis” at ISBI 2021 (M. Lorenzi), the Workshop on Disease Progression Modelling at MICCAI 2021 (M. Lorenzi), the international conference JURIX 2020 (S. Villata, Program Chair), the "Sister Conference Best Papers" track of IJCAI 2020 and IJCAI 2021 (S. Villata), 7th Workshop on Argument Mining at EMNLP 2020 (E. Cabrio, S. Villata), the Workshops Track of ACL 2022 (E. Cabrio), the IEEE ISBI conference 2021 (L. BlancFéraud), 6G Wireless foundations forum 2021 (D. Gesbert, General Chair), IPAM (UCLA) long program “Mathematical challenges and opportunities for Autonomous vehicles” (P. Goatin), and the Eurographics conference 2019 (P. Alliez, Program Chair).

Coordination and participation in international research projects The activity of the chairholders at the international level is highlighted by their involvement in international research projects both as coordinators and as participants. Among others, M. Filippone (partner of the EU international training network in Machine Learning for Communication Systems), C. Bouveyron, R. Flamary, and M. Gori (members of the European Lab for Learning and Intelligent Systems [ELLIS] society), S. Villata, and E. Cabrio (principal investigators of the EU CHIST-ERA ANTIDOTE project “Argumentation-driven explainable artificial intelligence for digital medicine”, 2021-2024), P. Barbry (partner of the H2020 DiscovAIR project on “Discovering the cellular landscape of the airways and the lung” on SARS-CoV-2 virus), D. Rouquié (partner of the European Research & Innovation project (H2020) EU RiskHunter), M. Onen (principal investigator of the Papaya European project “Platform for privacy preserving data analytics”), C. Bouveyron, E. Cabrio, F. Gandon, and S. Villata (3IA members in the EU Project AI4MEDIA, a Center of Excellence delivering next generation AI Research and Training at the service of Media, Society, and Democracy), and M. Sermesant (principal investigator of the EU H2020 Project SimCardioTest). Two Advanced ERC grants are currently ongoing in the Institute, held by the chairholders X. Pennec and R. Deriche.

Local integration projects Two actions have been selected here to show the potential of conjugating fundamental and experimental research across research Axis 1 and Axes 2, 3 and 4, respectively. •

The first project is FedBioMed led by M. Lorenzi (Axis 1) on the use of Federated Learning in a network of cancer hospitals in France (members of the UNICANCER consortium) and abroad (UK). This project also involves the chairholders O. Humbert (Axis 2), S. Ourselin (Axis 2,UK), and M. Onen (Axis 4).

•

The second project consists in the collection of the contributions of 3IA chairholders during the COVID-19 pandemic to help advance research on SARS-CoV-2 and related issues. These contributions resulted in the extraction of information from clinical text data and its enrichment through medical ontology concepts (F. Gandon, S. Villata, A. Tettamanzi, and E. Cabrio), ethical challenges in moderating disinformation during the pandemic (S. Villata for the National Committee for Digital Ethics), the quantification of the concentration of the virus and determining the presence of the different variants in collaboration with Nice Côte d’Azur Metropolis and Veolia (P. Barbry), the use of a statistical network clustering method (Linkage) to summarize Covid-19 scientific publications (C. Bouveyron), and a novel visualization tool boosting the analysis of protein interfaces to study commonalities between the spikes of SARS-Cov-1 and SARS-Cov-2 (F. Cazals).

Research activity of Axis 1 during the period 2019-2021

Research topics The objective of this axis is to develop fundamental AI models and algorithms for realworld problems. Within this framework, a core topic is naturally Machine Learning (ML) which ranges from statistical learning to deep learning and their theoretical aspects. Different models of supervised and unsupervised learning have been proposed by 3IA Côte d'Azur chairholders, notably for learning from heterogeneous data, for topological data analysis and for Natural Language Processing (NLP). Knowledge Representation and Reasoning (KRR) is also well represented in this axis, with approaches aiming at combining machine learning with symbolic methods, knowledge representation and processing on the Web, methods for linking unstructured, structured and semantic data, and finally reasoning on complex heterogeneous dynamic networks. A section of Axis 1 is dedicated to constraint-based AI, where distributed and federated AI techniques have been proposed in addition to formal methods for reasoning under uncertainty, especially for real-time decision-making in scenarios specific to the other three axes of 3IA. Finally, the current and crucial problem of developing AI methods that are interpretable and explainable is of fundamental importance in this context, together with the challenges of AI ethics, which is also one of the research topics pursued in this axis.

3IA Chairs A total of 19 principal researchers have contributed to Axis 1: •

12 chairholders since 2019 (J.D. Boissonnat, C. Bouveyron, F. Delarue, M. Filippone, R. Flamary, F. Gandon, M. Lorenzi, X. Pennec, J.C. Regin, C. Simpson, A. Tettamanzi, and S. Villata)

•

4 chairholders since 2021 (E. Cabrio, M. Kanagawa, P.-A. Mattei, and G. Neglia)

•

1 international chairholder (M. Gori, 2019)

•

2 affiliate chairholders (F. Limpens, 2020 and G. Ottosson, 2021)

International Chair - 2019

Marco Gori - University of Siena Learning and reasoning with constraints Learning and inference are traditionally regarded as the two opposite, yet complementary and puzzling components of intelligence. In the last few years, Prof. Gori has been carrying out research on constrained based models of the environmental agent interactions, with the main purpose of unifying learning, inference, and reasoning within the same mathematical framework. The unification is based on the abstract notion of constraint, which provides a representation of knowledge granules, gained from the interaction with the environment, as well as of supervised examples. The theory offers a natural bridge between the formalization of knowledge, expressed by logic formalisms, and the inductive acquisition of concepts from data.

Chairs holders - 2019

Jean-Daniel Boissonnat - Inria Topological data analysis We are studying the mathematical, statistical, algorithmic and applied aspects of topological data analysis, a fast-growing field with a well-funded theory that is attracting increasing interest in both fundamental research and in industry. Our ambition is to uncover, understand and exploit the topological and geometric structures underlying complex data.

Charles Bouveyron - Université Côte d'Azur Generative models for unsupervised and deep learning with complex data We focus on learning problems that are made difficult by real-world constraints, such as unsupervised deep learning, choosing a deep architecture for a given situation, learning from heterogeneous data or in ultra-high-dimensional scenarios. We seek to develop deep generative models, encoding sparsity priors, to address those issues.

François Delarue - Université Côte d'Azur Mean field multi-agent systems in AI

We study AI systems with a large number of rational agents with mean field interactions. Theoretical questions remain open, specifically when related Nash or Pareto equilibria are not unique, and thus corresponding numerical and learning methods are key issues. Applications include neural networks, power grids, crowd management, cybersecurity, etc.

Maurizio Filippone - Eurecom Probabilistic machine learning

Probabilistic machine learning offers a principled framework for quantification of uncertainty across various sciences. The Chair will tackle three major modeling and computational issues: (i) the need to develop practical and scalable tools for accurate quantification of uncertainty, (ii) the lack of interpretability, and (iii) the unsustainable trend in energy consumption.

Rémi Flamary - Université Côte d'Azur Optimal transport for machine learning

The main objective of this project is to change the way we learn from empirical data using optimal transport. We will first investigate optimal transport for transfer learning with biomedical and astronomical applications. Second, we will adapt the Gromov-Wasserstein distance for structured data and transfer between deep learning models with different architectures.

Fabien Gandon - Inria Combining artificial and augmented intelligence technics on and through the web Formalizing knowledge-based models and designing algorithms to manage interactions between different forms of artificial intelligence (e.g. rule-based, connectionist, and evolutionary) and natural intelligences (e.g. individual user, and crowd) on the web.

Marco Lorenzi - Inria Interpretability and security of statistical learning in healthcare Statistical learning in healthcare must ensure interpretability and compliance with secured data access. To tackle this problem, I will focus on 1) interpretable biomedical data modeling via probabilistic inference of dynamical systems, and 2) variational inference in federated learning for the modeling of multicentric brain imaging and genetics data.

Xavier Pennec - Inria Geometric statistics and geometric subspace learning We study the impact of topology (singularities) and geometry (non-linearity) of the data and model spaces on statistical learning, with applications to computational anatomy and the life sciences. The tenet is that geometry is critical when learning with limited resources and real-world constraints such as small data and limited computational resources.

Jean-Charles Régin - Université Côte d'Azur Decision intelligence

We are designing explainable decision-making processes satisfying real world constraints in a multi-objective environment including incomplete, fuzzy or stochastic data.

Carlos Simpson - CNRS AI and mathematics

My research addresses the interactions between research areas in algebra, category theory and geometry, and machine learning. This includes applications of AI to the classification of interesting algebraic and geometric structures, and the interactions between AI and formal verification of proofs in both directions.

Andrea G.B Tettamanzi - Université Côte d'Azur Towards an evolutionary epistemology of ontology learning I am developing symbolic learning methods based on evolutionary computation to overcome the knowledge acquisition bottleneck in knowledge base construction and enrichment. This project, straddling machine learning and knowledge representation and reasoning, combines symbolic aspects of AI with easily parallelizable computational methods.

Serena Villata - CNRS

Artificial argumentation for humans The goal of my research is to design and create intelligent machines with the ability to communicate with, collaborate with, and augment people more effectively. To achieve this challenging goal, intelligent machines need to understand human language, emotions, intentions, behaviors, interact at multiple scales, and be able to explain their decisions.

CHAIRS 2021

Elena Cabrio - Université Côte d'Azur AI and natural language The goal of my research is to design debating technologies for advanced decision support systems, to support the exchange of information and opinions in different domains (as healthcare and politics), leveraging interdisciplinarity and advances in machine learning for Natural Language Processing.

Motonobu Kanagawa - Eurecom Machine Learning for Computer Simulation

Computer simulation has been widely used for planning high-impact decision-making (e.g., policies on climate change and Covid-19), but its reliability depends on how accurate simulations can imitate reality. This project develops machine learning methods to improve a simulator’s reliability and the resulting decision-making.

Pierre-Alexandre Mattei - Inria Deep learning for dirty data: a statistical perspective The successes of machine learning remain limited to clean and curated data sets. By contrast, real-world data are generally much messier. We work on designing new machine learning models that can deal with “dirty” data sets that may contain missing values, anomalies, or may not be properly normalised. Collaborators include doctors and astronomers.

Giovanni Neglia - Inria Pervasive Sustainable Learning Systems (PERUSALS) PERUSALS (Pervasive Sustainable Learning Systems) seeks to identify design principles of Internet-scale distributed learning systems, with a focus on the tradeoff between performance (in particular training and inference times), economic and environmental costs, and privacy.

3IA AFFILIATE CHAIRS 2020

Freddy Limpens - Inria | MNEMOTIX Big Knowledge graphs evolution and life cycle When it comes to deploying big Knowledge Graphs (KG) based information systems in the real world, be it for distributed organisations, small or big companies, some fundamental problems remain unanswered. Our research focuses on solving some of them such as managing the history and evolution of big KG, offering rapid access to KG's data, automatizing the validation of data, and providing generic templating language for more easily exploiting query results.

2021

Greger Ottosson - Inria | IBM Trustworthy AI and Explainable Decisions for Business Automation As we apply Machine Learning to automate decisions in financial services, healthcare and government, there is increasing user need and regulatory demand for transparency and explainability. Our AI research is focused on explainability for decisions that combine ML-based predictions and rule-based business policies.

Actions and highlights •

Geometric statistics and topological data analysis: Key algorithmic aspects, such as new approaches to compute persistent homology and a manifold tracing algorithm in arbitrary dimensions are studied to uncover and exploit the topological and geometric structure underlying complex and possibly high-dimensional data, leading to very efficient implementations (Chair: J.-D. Boissonnat). The foundations of statistics in manifolds are investigated to understand how the statistical estimations diverge from the Euclidean case with curvature, singularity, and stratification (Chair: X. Pennec).

•

Supervised and unsupervised deep learning: Deep generative models are developed to address unsupervised deep learning from heterogeneous or ultra-high-dimensional data, e.g., pharmacovigilance data. A dynamic extension of the statistical model called STBM is proposed to cluster the nodes of a dynamic graph, accounting for the textual edges and their frequency (Chair: C. Bouveyron). Always in the context of unsupervised deep learning, the definition of a constraint-based modeling of the environment is examined to unify learning and inference within the same mathematical framework (Chair: M. Gori). Optimal transport (Chair: R. Flamary) to provide a novel formulation and modeling on graph data is also inspected, and applied to the generation of adversarial samples on remote sensing images to promote robustness. Furthermore, research investigates how to design new ML models that can deal with “dirty” data sets, containing missing values, anomalies, or not properly normalized (Chair: P.-A. Mattei).

•

AI and mathematics: The mean field paradigm describes large interacting particle systems (e.g., neurons or robots) through their sole common distribution. Both the cooperative (mean field control problems) and the competitive (mean field games) scenarios are investigated casting them as stochastic control problems/games, but for an infinite population of agents (Chair: F. Delarue). Also, the proof construction and their formal verification are investigated, such as the utilization of reinforcement learning to learn how best to choose the steps in a classification proof for finite algebraic structures (Chair: C. Simpson).

•

Bayesian deep learning and Gaussian processes: A novel approach to specify sensible priors is proposed both for supervised (regression, classification) and unsupervised learning (generative models, data compression), to accurately quantify uncertainty in model parameters and predictions. Also, novel ways are proposed to initialize the optimization based on Bayesian linear models (Chair: M. Filippone).

The development of a machine learning methodology for calibrating and validating a computer simulator is investigated to improve the simulator's reliability and the resulting decision-making, with applications in pension reform and tsunami early warnings (Chair: M. Kanagawa). •

Interpretability, explainability, sustainability and security in AI: Interpretable data modeling approaches for time series of heterogeneous high-dimensional data are investigated to explain automated decisions in financial services, healthcare, and government (Chair: G. Ottosson, IBM). Also, federated learning methods (Fed-BioMed) for the modeling of multi-centric data with various degrees of heterogeneity (across clients’ features and labels) are proposed and successfully deployed in different hospitals in France (Chair: M. Lorenzi). Given the urgent need to find a satisfiable trade-off between performance in ML (in particular training and inference times) and the huge economic and environmental costs, a study examines how to better configure existing machine learning systems, and how to design the next-generation system (Chair: G. Neglia).

•

Knowledge representation and reasoning: Both formal and empirical methods are proposed to couple, hybridize or coordinate different forms of intelligence, i.e., artificial and natural intelligences on the Web, by relying on ontology-based representations and Semantic Web formalisms, e.g., different strategies for knowledge sharing in open multi-agent systems (Chair: F. Gandon). Knowledge engineering problems are also investigated from an industrial perspective, focusing on the life cycles of big graphs (Chair: F. Limpens, Mnemotix). Considering the challenge and expense of checking the consistency and inferring implicit knowledge from knowledge graphs, an evolutionary algorithm is proposed to efficiently discover subsumption, class disjointness axioms, involving atomic and complex class expressions (Chair: A. Tettamanzi). Furthermore, novel algorithms and data structures are proposed for solving multi-objective and multi-scale problems such as multimodal multi-objective transport route searches (Chair: J.-C. Regin).

•

Natural language processing: New methods are investigated to leverage advances in ML for NLP, notably methods targeting debating technologies and subtle and implicit language. New natural language argument extraction, and classification and generation methods are investigated more particularly, with a focus on legal and medical texts, political debates, and harmful content on social media, e.g., hate speech, disinformation, and propaganda (Chairs: S. Villata, E. Cabrio).

Research activity of Axis 2 during the period 2019-2021 Research topics This axis aims to generate new fundamental knowledge in medicine to create computational models of the human body and to analyze digital health data in order to improve medical practice in disease prevention, diagnosis, therapy monitoring, and patient rehabilitation. To achieve this purpose, this axis uses data-driven methodological approaches based on machine learning, but also on the integration of mathematical models of human anatomy and physiology. A typical example is the generation of a digital patient which integrates patient data of various natures and at various temporal and spatial scales and analyzes these data to solve a wide range of clinical tasks. A major issue for this axis is how to manage personal and health data while enforcing data privacy and security requirements. This can be handled both by the development of new technologies (such as federated learning) and by the action of dedicated organizations (hospitals, health data hubs, etc.). The main clinical applications of this axis are in neurology, neuroimaging, oncology (prostate, lung, thyroid, abdomen, etc.), cardiology, vascular diseases, nuclear medicine, digital pathology, and functional rehabilitation.

3IA Chairs A total of 11 researchers have contributed to Axis 2: •

7 Chairs since Oct. 2019 (N. Ayache, H. Delingette, M. Sermesant, R. Deriche, J-P. Merlet, O. Humbert, and F. Bremond),

•

3 Chairs since Sept. 1, 2021 (V. Zarzoso, J. Raffort, M. Zuluaga),

•

1 International chair (S. Ourselin) since 2019.

Two Chairs are Medical Doctors (O. Humbert, J. Raffort).

International chair - since 2019

Sébastien Ourselin - King’s College London AI for Healthcare & Image-guided Interventions. He is Head of School of Biomedical Engineering & Imaging Sciences, Professor at King’s College London and Director of the EPSRC Image-Guided Therapies UK Network+. He published more than 400 articles (h-index: 64). He is associate editor for IEEE Tr. on Med Imaging,J of Med Imaging, Nature Scient. Reports, and Medical Image Analysis.

Chairs 2019

Nicholas Ayache - Inria

AI for e-patients and e-medicine We are designing and exploiting modern AI methods: (i) to personalize the parameters of advanced models of the e-patient, (ii) to drive e-medicine algorithms on personalized e-patients (i.e. digital twins) for automated diagnosis, prognosis and therapy, in an efficient, robust, safe and explainable manner. We also seek to augment databases with biophysical simulation.

François Brémond - Inria

Video analytics for human behavior understanding Video analytics enables us to measure objectively the behavior of humans by recognizing their everyday activities, their emotion, eating habits and lifestyle. Human behavior can be modeled by learning from a large quantity of data from a variety of sensors to improve and optimize, for instance, the quality of life of people suffering from behavior disorders.

Hervé Delingette - Inria

Joint biological and imaging biomarkers in oncology We exploit joint information from imaging and biological data to improve the diagnosis and treatment planning, focusing on lung cancer. This approach relies on methods involving unsupervised deep learning, uncertainty quantification, sparse Bayesian feature selection and the handling of confounding factors.

Olivier Humbert - Université Côte d'Azur

Comprehensive omics profiling for precision medicine in Oncology I am combining various patient extracted “omics” data, including multimodal imaging features, for integrative and data-driven computational medicine. I focus on challenging fields in oncology such as (i) radiogenomics and outcome-focused research in metastatic breast cancer and (ii) the accurate prediction of response to immunotherapy.

Jean-Pierre Merlet - Inria

Non-invasive assessment of disabilities We use mathematical/AI methods for (i) designing non-intrusive and affordable monitoring/assistance devices that are adaptable to the user’s/doctor’s needs, (ii) deducing medically pertinent health-indicators from the data, taking into account measurement errors, and (iii) detecting rare events that may be the sign of emerging pathology.

Maxime Sermesant- Inria

AI and biophysical models for computational cardiology The application of AI in healthcare is challenging due to its lack of robustness and explainability. This project aims to introduce physiological priors in AI through biophysical models. This can be done by reformulating problems through such models, by learning spatiotemporal dynamics from biophysics or by augmenting features and data with such simulations.

Rachid Deriche - Inria Computational brain connectomics This project will reconstruct and analyze the brain’s neural connections network, the connectome, via a computational brain connectomics framework based on ground-breaking AI algorithms and ML tools to gain insight into brain architecture, functioning and neurodegenerative diseases.

Chairs 2021

Juliette Raffort-Lareyre - CHU Nice / Université Côte d'Azur Applications of AI for patients with vascular diseases Our team develops AI-derived applications for patients with vascular diseases including aortic aneurysm. We aim to create aid-decision support systems to enhance evidence-based decision and precision medicine through a translational approach including: identification of biomarkers, automatization of vascular imaging analysis, development of predictive scores using machine learning.

Vicente Zarzoso - Université Côte d'Azur IAblation: Artificial Intelligence for patient-centered atrial fibrillation ablation Atrial fibrillation is the most common sustained arrhythmia in clinical practice and remains the last great frontier of cardiac electrophysiology. The project aims to put forward new AI techniques to help cardiologists perform more effective patient-tailored catheter ablation procedures to treat this challenging cardiac condition.

Maria A.Zuluaga - Eurecom

Learning-based Models in Medical Imaging: Closing the Gap towards Clinical Translation We aim to close the gap hindering the translation of AI systems into clinical practice. To this end, we develop novel AI tools for healthcare that 1) learn incrementally as data become available, 2) generalize across multi-modal data and 3) perform quality control of a model and can generate alerts when a decreased in performance is detected.

Actions and highlights •

Federated learning: The Fed-BioMed project led by M. Lorenzi (Axis 1) is a nationwide initiative that is currently being tested and that concerns the analysis of PET images (Chair: O. Humbert) in real-world conditions using the infrastructure of a local hospital. The extension of this framework to large-scale international clinical projects is actively studied by the international chairholder, S. Ourselin.

•

Medical imaging analysis and multiomics: Medical images are processed by deep learning algorithms to detect abdominal organs in ultrasound images, to create biomarkers for immunotherapy patient selection from whole slide imaging, and to detect cancer lesions in MR prostate imaging (Chairs: N. Ayache and H. Delingette). The analysis of medical images is extended to include biological (omics) data to create more powerful biomarkers and to develop precision medicine. This meta-analysis is performed in the context of lung cancer by combining radiomics and metabolomics (Chair: O. Humbert) but also by combining radiomics and transcriptomics to improve the treatment of patients with aortic aneurysm (Chair: J. Raffort). For these tasks, a major concern is to create high-quality datasets and to monitor the performance of algorithms across heterogeneous and unseen datasets (Chair: M. Zuluaga).

•

Cardiac imaging and signals: AI algorithms are developed to create patient-specific electromechanical models of the heart but also to create physics-informed networks capable of simulating cardiac electrophysiology (Chair: M. Sermesant). Furthermore, data-driven approaches are investigated to perform patient selection and therapy guidance in the context of catheter ablation for the treatment of atrial fibrillation (Chair: V. Zarzoso).

•

Brain connectomics: The relation between brain connectivity and neurological diseases is investigated through the generation of novel biomarkers partially based on diffusion MRI (Chair: R. Deriche).

•

Patient monitoring: The video-based monitoring of the everyday activities of elderly subjects in real-world settings was developed to produce new reference datasets for the field and to include the detection of emotions (Chair: F. Bremond). Furthermore, ML was used to improve the control of a cable-driven parallel robot that can assist frail patients to stand and walk (Chair: J-P. Merlet).

Research activity of Axis 3 during the period 2019-2021

Research topics Axis 3 aims at quantifying biological structures in space and time from new-generation data at all scales (molecules, cells, tissues, organs) in order to understand and model biological phenomena. Data are heterogeneous, massive, high-dimensional, and embed complex correlations. Novel AI technologies are required to characterize normal and pathological biological functions and to understand the adaptation of organisms to their changing environment (computational biology). Conversely, insights from cognition offer exciting perspectives to propose new bio-inspired AI algorithms and architectures (bio-inspired AI).

3IA Chairs A total of 8 researchers have contributed to Axis 3: •

4 Chairs since Oct. 2019 (L. Blanc-Féraud, F. Cazals, P. Reynaud-Bouret, E. Van Obberghen-Schilling)

•

2 Chairs since Sept. 1, 2021 (P. Barbry, B. Miramond)

•

1 International Chair (D. Wales)

•

1 Affiliate Chair (D. Rouquié) since 2020

International chair - since 2019

David Wales - University of Cambridge Solution landscapes for Machine Learning

We explore machine learning landscapes in the cost function parameter space, which isanalogous to the potential energy surface of a molecule as a function of atomic coordinates. Ongoing advances in methodology developed in chemical physics, can therefore be immediately applied to ML solution landscapes. Our objectives are to use these tools to design improved predictions, and apply them to problems in molecular science and health care. In particular, we seek improved machine learning tools for clinician diagnostic support, to provide earlier detection of the deteriorating (and improving) patient. Specific applications include prediction of readmission to intensive care, which represent a failure in down-transfer to the ward, and are often associated with patient mortality.

Chairs 2019

Laure Blanc-Féraud - CNRS Imaging for biology

Recent advances in microscope technology provide outstanding images that allow biologists to address fundamental questions. This project aims at developing new AI methods and algorithms for (i) novel acquisition setups for super resolution imaging, and (ii) extraction of valuable quantitative information from these large heterogeneous datasets.

Frédéric Cazals - Inria

AIMS: Artificial intelligence for molecular studies By learning essential features of proteins and their complexes, we shall deliver biologically relevant information for large molecular systems on biologically relevant time scales, leveraging our understanding of biological functions at the atomic level, and providing key inputs for protein design and engineering, and protein interaction networks.

Patricia Reynaud-Bouret - CNRS MEL: Modeling and estimating learning

We are defining new probabilistic models and new estimation methods to understand the deformation of functional connectivity during learning in in vivo experiments.

Ellen Van Obberghen-Schilling - Inserm AI-powered analysis of the tumor microenvironment

Our project will integrate tissue imaging modalities and artificial intelligence-based analysis tools for a deeper understanding and control of cancer, targeting tumor microenvironment and on the role of the extracellular matrix (ECM) in carcinoma progression, spread and response to therapy.

Chairs 2020

Pascal Barbry - CNRS Human Lung Atlas

The project elaborates on state-of-the-art approaches in genomics and cell biology to describe complex biological samples at the single-cell resolution. Multidimensional biological experiments result in large scale descriptions of DNA, RNA and protein expressions that can be integrated in time and space. The project aims at: (1) developing novel data-mining approaches based on machine learning and AI; (2) apply them to the study of the normal and pathological lung, in the context of serious threats that touch this organ (COVID-19, asthma, cystic fibrosis, cancer,...).

Benoît Miramond - Université Côte d'Azur

Bio inspired AI from neurosciences to embedded autonomous devices The research project seeks to draw on the structure and function of the biological brain to develop more energy-efficient AI methods and algorithms. The scientific approach ranges from neural dynamics to the emerging cognitive properties of these networks and ultimately to the design of embedded neuromorphic electronic circuits. The project will focus on building bridges between the NeuroMod neuroscience institute and the 3IA Cote d'Azur institute.

Affiliate Chair 2020

David Rouquié - Université Côte d'Azur | BAYER Human health chemical risk assessments

The main focus of my research activity as affiliate chair at 3IA Côte d’Azur is on the emerging theme of chemical safety by design. This theme derives from all the on-going initiatives to improve the characterization of the risk of chemicals to humans by using more data from non-animal technologies and goes far beyond. Indeed, we are living a paradigm shift in the way bioactive small molecules are discovered but also de-risked. Instead of relying on numerous, long, costly cycles of trials and errors, thanks to the advance in systems biology and state-of-the-art machine learning algorithms it is possible to proactively drive de novo chemical design with high probability to induce specific biological responses. For the first time, we have shown as proof of concept that a learning procedure can automatically design molecules that have a high probability to induce a desired transcriptomic profile in cell lines. In my position of affiliate chair at 3IA Côte d’Azur, this approach will be further developed by building the pillars necessary to guide chemical design toward optimized safety profiles while maintaining the desired biological effect of the compounds.

Actions and highlights •

Single cell: Characterization of the individual cells of lungs and airways contributes to research on the SARS-CoV-2 virus and the Human Cell Atlas. New methods are proposed to analyze the structure of RNAs at single-cell resolution and to obtain the spatial transcriptomic data needed to detail the mutational landscape in tumors (Chair: P. Barbry).

•

Molecular studies: The essential features of proteins discovered by statistical analysis or by combining sequence-based and structural data analyses can reveal biological functions at the atomic level. This contributes to a better understanding of interfaces of proteins within complexes and their dynamics, and provides key information about protein design and engineering (Chair: F. Cazals). Theoretical results have also been obtained for the characterization of the energy landscape used to support the clinician’s early diagnosis using electronic patient health data (Chair: D. Wales). Also, chemical-safe molecules with the desired biological responses are designed from transcriptomic and cell-painting data using GAN (Chair: D. Rouquié, Bayer).

•

Super-resolution microscopy and extracellular matrix: This research exploits the recent results obtained in sparse optimization and combines them with the relevant acquisition protocols to obtain 2D and 3D super-resolution optical microscopy images of living cells, allowing the study of fine structures in and outside the cells (Chair: L. Blanc-Féraud). Also, structural biomarkers extracted from ECM network images in vitro of the tumor microenvironment have provided useful information for the study of tumor development, prognosis and response to treatment, in addition to standard cell analyses (Chair: E. Van Obberghen-Schilling).

•

Bio-inspired AI: This topic is studied both from a behavioral and neuronal point of view. It states complex theoretical questions using spike sorting, simulation of Hawkes processes, mean-field limits or reconstruction of functional connectivity. This last method has been proven to be able to follow the mnesic trace and investigate where a memory is encoded (Chair: P. Bouret). Furthermore, the development of a spike neural network architecture or a self-organizing neural architecture inspired by the brain leads to low-power classification algorithms (Chair: B. Miramond). Realizations are embedded in autonomous devices, such as the VS23 Soyuz space flight or in smart glasses.

Research activity of Axis 4 during the period 2019-2021

Research topics The concept of smart and secure territories refers to sustainable cities, non-urban areas and communities with networked intelligent systems that place the users at their core and deliver personalized services and resources suited to a wide range of behaviors, constraints, and preferences. Such intelligent systems should help sense complex situations, make decisions, predict risks, extract knowledge from data, and report back to users. In addition, they should address the challenges of secure territories: reliability, safety, security, legality, trust, and acceptability.

3IA Chairs A total of 8 principal researchers have contributed to Axis 4: •

3 Chairs since 2019 (M. Teller, M. Önen and P. Alliez)

•

4 Chairs since 2020 (E. Di Bernardino, D. Gesbert, P. Goatin, C. Richard)

•

1 Affiliate Chair (A. Lhéritier) since 2020.

In addition, this axis leverages the research results developed in Axis 1.

Chairs 2019

Pierre Alliez - Inria

3D modeling of large-scale environments for the smart territory We are exploring the generation of rich 3D vector maps with semantic attributes from raw measurement data. We plan to learn geometric priors and error metrics that locally adapt to the semantic class of objects. We are developing a pliant approach with the capability to model the wide range of objects, which abound in open environments of the smart territories.

Melek Önen - Eurecom

Privacy-preserving machine learning Machine learning has become popular due to cloud computing technology and the increasing number of datasets. Outsourcing computations poses a risk to data privacy. Therefore, our goal is to explore privacy-preserving variants of machine learning techniques while leveraging novel cryptographic methods.

Marina Teller - Université Côte d'Azur Deep law for tech (DL4T)

The DL4T team is building the legal framework for deep technologies. Starting from the observation that law is often perceived as a force of resistance to innovation, we want to position our research upstream of technology, to support the emergence of technical standards and to promote a convergence between law and AI.

CHAIRS 2020

Elena Di Bernardino - Université Côte d'Azur

Territorial Security through environmental risks management This project deals with risk assessments related to environmental extreme events. Analyses and predictions of floods, summer heatwaves, and storms are significant questions facing statisticians and risk assessors. Such environmental risks are the result of a long chain of casualties, involving several aleas, often correlated, with complex spatio-temporal dependent structures among extremes. Our contributions in the prevention and management of environmental risks, will be twofold: 1/ Proposing novel and realistic definitions of risks indicators in environmental contexts.2/ Studying in-depth their statistical inference, i.e. specifying more accurately the associated uncertainties. In this project, the skills required to handle the modeling of these uncertainties are stochastic processes and random fields, spatio-temporal models, multivariate extreme theory, as well as practical expertise on spatial and environmental data gathered from firms in 3IACôte d’Azur.

David Gesbert - Eurecom

Internet of Learning Thing, a machine learning approach to future IoT networks In this chair, we develop cooperative forms of decision making, that can be implemented on distributed IoT devices, and not relying on the assumption that all data is centralized in the cloud. IoT devices can learn to coordinate with each other in their usage of the wireless spectrum, energy, and other resources while dealing with arbitrary noise uncertainties in their observation data. Cooperative machine learning will bring a profound evolution in IoT system design, both at the level of radio access, as well as in the manner services will be orchestrated and how resources will be allocated.

Paola Goatin - Inria

Data driven traffic management This project aims at contributing to the transition to intelligent mobility management practices through an efficient use of available resources and information, fostering data collection and provision. We focus on improving traffic flow on road networks by using advanced mathematical models and statistical techniques leveraging the information recovered by real data. We are committed in creating a network of local stakeholders sharing knowledge and expertise.

Cédric Richard - Université Côte d'Azur

Distributed dark fiber optic sensing for smart cities monitoring Optical fiber, in addition to being a means of transmitting information, is also a material that is very sensitive to environmental variations. When a laser light pulse travels through an optical fiber, it interacts with tiny impurities in the material and optical backscattering occurs. The round-trip time of the light provides the locations of interactions and allows us to infer a backscattering profile along the fiber. Processing this response provides estimates of the local variations in temperature, deformation or acoustic pressure along the fiber. This technique, called Distributed Fiber Optic Sensing (DFOS), is currently experiencing growing interest. The goal of our project is to develop a breakthrough framework for smart cities monitoring based on DFOS over existing dark fibers, and Artificial Intelligence.

Affiliate chair - 2020

Alix Lhéritier - Eurecom | AMADEUS

Improving the Air Travel Experience via Probabilistic Regression with Epistemic Uncertainty Estimation in Adversarial Scenarios Alix Lhéritier was born in Montevideo, Uruguay, in 1978. He received the Computer Engineer degree and the M.Sc. degree in computer science from the Universidad de la República, Montevideo, Uruguay, in 2004 and 2010, respectively. In 2006, he was a Visiting Research Scholar at the Mathematical Sciences Research Institute, Berkeley, CA and at the Image Processing Laboratory of the University of Minnesota, MN. In 2011, he joined the Algorithms-Biology-Structure (ABS) team at Inria Sophia Antipolis, France as a doctoral student and he received his Ph.D. in computer science from the Université Nice Sophia Antipolis in 2015. He is currently with Amadeus, France, working on machine learning research and applications for the travel industry. His research interests include sequential decision problems, statistical dissimilarity and choice modeling.

Actions and highlights: •

AI & Law: The fast-paced development of AI makes it challenging to regulate, and to develop laws and ethical principles for questions pertaining to AI systems, connected objects, and digital twins. AI methods offer novel opportunities for devising tools capable of improving the quality and efficacy of justice, and contributing to a future era of deep law, digital evidence and predictive justice (Chair: M. Teller). The topics of AI ethics and “ethics by design” are investigated by S. Villata (Axis 1) and 3IA Assistant Professor A. Nogales.

•

Secure territories: Privacy and security concerns are raised by the storage, transmission, and processing of sensitive data by ML services. Novel cryptographic methods have been used to devise privacy-preserving neural networks, and robust watermarking solutions have been extended to neural networks (Chair: M. Önen). Privacy issues are also investigated in the medical data domain, in collaboration with M. Lorenzi (Axis 1) and O. Humbert (Axis 2). Furthermore, ML is used to improve the design of wireless networks for IoT via a federated machine learning approach that provides efficient beam tracking. Novel wireless networks, beyond 5G, are devised to deliver machine learning services (Chair: D. Gesbert).

•

Smart territories: Unused telecommunication fibers are used as intelligent distributed sensors for predicting earthquakes and monitoring road traffic. Contributions have been made to deep deconvolution and self-supervised learning from acoustic sensing data (Chair: C. Richard). The prevention and modeling of environmental risks such as forest fires is also tackled through spatiotemporal analysis of massive geospatial data recorded over decades. Methods such as statistical inference in large dimension, stochastic processes, multivariate extreme theory and data assimilation are used to model fire dynamics and devise novel weather-informed risk indicators (Chair: E. Di Bernardino). Another sensitive issue deals with the improvement of traffic flow on road networks, which is obtained through the analysis of real-world data, and through multi-scale mathematical modeling and prediction leveraging a novel Bayesian uncertainty quantification method (Chair: P. Goatin). Finally, contributions have been made to the analysis of satellite images (remote or onboard via low-power neural networks), to the automated generation of vector maps from satellite images, and to the physics-informed geometric reduction of 3D models for the radiative thermal simulation of satellites (Chair: P. Alliez).

Conclusion on the Research Activity during the period 2019-2021 To conclude, through the dynamism of the chairholders and their teams, the Institute has achieved significant results both at the international level (with the organization of scientific events held in major venues for AI and ML, their publication activity and the keynote talks they have been invited to deliver) and at the national level, where it has strengthened ties with local organizations such as the Nice Côte d’Azur Metropolis or the National Park of Mercantour, and hospitals such as the Nice University Hospital and the Antoine Lacassagne Center (CAL). These results are grounded on high-quality scientific contributions to the core elements of AI and their applications in computational medicine and biology, and smart territories. These applications show the concrete impact of AI methods and technologies on society, highlighting its beneficial contribution to the improvement of two main aspects, i.e., AI-assisted healthcare, and smart cities and territories, while always remaining cautious with respect to the related issues of security, confidentiality, and ethics (e.g., three chairholders (Ayache, Ourselin, Teller) are members of the Scientific Advisory Committee of the French Health Data Hub, and S. Villata is a member of the National Committee for Digital Ethics.

EDUCATION AND TRAINING

Design and strengthen the AI training offer for beginners and specialists 3IA Côte d’Azur offers a large number of training programs with the aim of increasing the number of students and professionals with AI skills and ensuring that they are able to meet the needs of companies. Profits generated by professional training, representing €160k for 2021, will be reinvested in training activities.

The 3IA label

was created to recognize 20 programs with an IA content >50%.

These educational programs include 5 master of science degrees (100% in English), 6 engineering degrees, 4 master’s degrees, 1 bachelor’s degree, 1 summer school, and

microcredential

courses.

In 2 years, the number of persons trained in AI (mainly in the fields of fundamental AI and applied AI) increased by 50%. Due to a dynamic communication policy, 3IA programs have gained an increased international

visibility,

attracting

students from all over the world. Geographical origin of the 3IA Students increased

The Institute human resources: +65 people in 3 years To supplement our permanent academic task force, we have asked our 44 PhD students to teach 32 hours of AI per semester. This teaching load was included in their contracts. From the beginning of 2019 until 2021, the teaching potential of our PhD students increased from 640 hours (1st academic year), 1,408 hours (2nd academic year), to 2,560 hours (current academic year). In addition, professors holding a 3IA Chair have taught a total of 68 courses since 2019, amounting to 4,075 hours. 75% of professors who hold a 3IA Chair and 3IA PhD students have taught their courses at the graduate level and 25% at the undergraduate level. Furthermore, we have strengthened our team by hiring a lecturer and 2 teaching engineers.

AI for specialists + 1100 people/year Specific high-quality innovative training courses have been designed for specialists in order to ensure that a sufficient number of them have frontier knowledge and competencies in fundamental and applied AI.

AI minor in 4 Graduate Schools Two AI minor courses taught in English (introduction and advanced) are offered in 4 Graduate Schools (master’s program open to PhD students). This represents 24 hours per semester, including 12 hours of lectures and 12 hours of field-relevant tutorials, all conducted by 3IA PhD students (totaling 72 hours). These AI minor courses are also open to professionals (lifelong training) and will be offered on the French Connected Campus (Campus connectés developed by the Ministry).

International : AI training shape the challenges of tomorrow

72 hours conducted by 3IA PhD students in 4 Graduate Schools

Ulysseus Within the European University Ulysseus, collaboration with the universities of Seville and Haaga Helia, both leaders in AI, will strengthen

the

Insti-

tute activities at the

A master’s degree program in collaboration with the Université Laval (Quebec)

European level. We are

The Institute launched a master’s degree pro-

this project is led by Haaga Helia and 3IA.

developing a master's program in AI with Ulysseus, the European ecosystem, and

gram in collaboration with the Université Lapartner of UCA and 3IA. This project also aims

“Artificial Intelligence Doctoral Academy” (AIDA)

at developing partnerships with Quebec com-

3IA Côte d´Azur also participates in the “Arti-

panies within the framework of internships.

ficial Intelligence Doctoral Academy” (AIDA)

val (Quebec), a long-standing and strategic

program, and is striving to become a European Other programs are being developed in part-

and world reference in postgraduate AI teach-

nership/collaboration with University of Bolo-

ing for PhD students (online courses in English).

gna, Italy and Santander University, Colombia.

The Deep Learning School The

Deep

gether

Learning

than

School 250

co-organized

engineers,

250

the

Institute

researchers

and

brings

to-

internation-

al PhD students for a week of training led by international experts.

"Ecole de l’Air” French Military Air School The Institute and the “Ecole de l’Air” French Military Air School have joined forces to develop not only AI initiation programs, but also AI training in space security with the highest degree of excellence.

AI for non-specialist students +1,200 people / year

AI modules have been integrated into existing programs in different fields and levels. Raising awareness of AI possibilities and tools is an important aspect of dissemination to students. Last year, 3IA launched an undergraduate specialty course in AI “Science and Technology” and a new AI track in the IT master’s degree.

An undergraduate specialty course in IA “Science and Technology” and a new AI track in the IT master’s degree.

The bonus points for volunteering in AI workshops are renewed each semester for all students of Université Côté d’Azur and the MIA (Maison de l’Intelligence Artificielle): 20 hours per semester/student.

A certification module in defense and cybersecurity The Law Graduate School created a certification module in defense and cybersecurity (Axis 4) open to a wide audience: it is currently offered as part of the IoT program and “Banking Law & FinTech” master’s degree. The objective is also to develop a micro-credential program for professionals. The Institute is currently working on creating a new minor course for non-specialists soon to be offered in 3 other Graduate Schools (School of Economics, Law, and Social Sciences).

AI for continuing education in companies +300 people/year Training modules for companies that combine space data, AI and entrepreneurship.

Customized training for STMicroelectronics

The Institute has been working with Safe Cluster

ment was signed with STMicroelectronics to train

(a competitiveness center) on a “Copernicus” Eu-

engineers and data analysts in AI. Face-to-face

ropean project with the support of CNES (French

training includes three five-day modules based on

National Center of Space Studies). By sharing our

the company’s study cases. The customized train-

expertise with professionals, we have identified an

ing program features lectures in the morning and

opportunity to create 4 new training modules for

tutorials in the afternoon using individual use cas-

companies that combine space data, AI and en-

es. This regional AI training project is supported by

trepreneurship. This project is developed by the

Université Côte d’Azur and Marseille University.

As an example of customized training, an agree-

chairholder, Pierre Alliez, and his PhD students, in collaboration with Université Laval. In the future, Amadeus and Sanofi are expected to sign the same kind of agreement.

Graduate training program “AI & Health” The Institute launched a graduate training program “AI & Health” in collaboration with the Nice University Hospital and is currently training 25 doctors from Southern France.

25 doctors trained

AI for citizen outreach +4,000 people/year

Lectures designed for a wide audience were organized to give all citizens the opportunity to understand how the society in which they live is evolving. EUROPIA invited 3IA chairholders to hold conferences for a wide audience on AI & Women (M. Teller), AI & Cybersecurity (C. Bouveyron), and AI & Ethics (S. Villata, M. Teller). The following indicators show the number of outreach activities carried out in collaboration with the Maison de l’Intelligence Artificielle (MIA) since October 2020: • school visits >3000 students • conferences and events for companies and regional authorities >520 people • and for a wide audience > 1,800 people (including online views)

Strategic development In the future, the Institute will continue to significantly increase the number of training programs for the different categories above. To maximize the human resources dedicated to our ambitious program, we will coordinate the major stakeholders in AI both at the regional and national levels. New partnerships with a local French Entrepreneur association (UPE06) and Skema Business School are being developed to create a certificate in management and AI. Regarding the HR bottleneck for teaching AI, the Institute’s dynamic led the Members to plan new recruitments in AI for the coming years.

AI and quantum technologies At the national level, AI and quantum technologies will be promoted in 2022 as quantum technologies will become a national priority in France.

“France 2030” The new French AI curriculum Furthermore, at the request of the French Ministry of Higher Education and Research, as part of the “France 2030” plan, the network of four 3IA Institutes will prepare, together with Inria, the new French AI curriculum, based on their training programs. This may in particular yield a national 3IA certification delivered by the 3IA Institutes, ensuring the quality and the coordination of the training offer.

PARTNERSHIPS AND INNOVATION

Bring scientific and technological added value to companies through collaboration agreements targeting their economic development The establishment of the Institute has accelerated the dynamic growth in AI fields throughout the French Riviera. Concretely, the Institute brings scientific and technological added value to companies through collaboration agreements targeting their economic development. It also contributes significantly to the development of the regional policy of Université Côte d’Azur and its partners in the innovation and partnership axis. To accelerate the generation of leads and opportunities, the Institute maintains continuous exchanges with the local company clusters ICAIR and Cluster IA.

Position research and technology actions in line with local policies Regarding Axes 2 and 4, the Institute’s major interactions are with the local authorities and public institutions, in order to position research and technology actions in line with local policies. Specifically, the Institute is developing partnerships with the Nice Côte d’Azur Metropolis, internationally visible in the field of Smart Cities, the Sophia Antipolis Urban Community, the largest European Technopole, and with the Alpes-Maritimes Departmental Council in its Smart Deal. To date, the Institute has undertaken 5 collaborative projects with the Nice Côte d’Azur Metropolis and the Alpes-Maritimes Department to scientifically characterize and use the data they have, especially in the fields of Smart Cities and Health.

Attract international business The Institute also contributes to the international attractiveness of the region by collaborating with the Team Côte d´Azur (TCA) agency. This agency, mandated by the French authorities and the local economic partners, works on attracting external businesses to settle in the Côte d’Azur. To date, the Institute has initiated discussions with about 20 companies introduced by TCA.

Socio-economic impact / INTERNATIONAL MANAGEMENT From the outset, the Institute established a “Partnership and Innovation Coordination Unit” as the point of contact to manage external requests and questions. This Unit brings together representatives of the Members’ technology development and transfer services, in liaison with their respective legal and financial services. It manages the various opportunities and monitors the performance indicators of the Institute, supported by a CRM.

/ STATUS ON PARTNERSHIPS WITH COMPANIES Despite the difficult current context, the business sectors remained

140 companies which the Institute is in contact with

dynamic, in particular those of health and the environment. The suspension of the R&D activities of certain companies that initially took part was compensated by new companies which started exchanges or collaborations with the Institute. As of today, there are more than

57 contracts signed

140 companies with which the Institute is in contact for actual or planned collaborations. A total of 57 contracts have been signed

€4.27M of cash income

between a company and one of the Members involving a researcher of the Institute and on a subject related to the themes of the Institute. Financially, these contracts represent €4.27M of cash income for

€10.8M of financial investment from companies

the Members. The overall financial investment of companies in these contracts, when including their own costs, amounts to more than €10.8M.

ContractsTypes Types Contract

Contracts per Axis

2% 14%

30%

39%

61% 2%

45% 2% Axis 1

Axis 2

Axis 3

Axis 4

CIFRE Thesis

Expertise

IP Licensing

IP Right transfert

Research

Post-Doc

As illustrated in the left figure above, about 60% of the companies are SMEs or startups. The figure on the right shows the distribution of contract types, indicating the strong interest of the companies for scientific collaborations, and the good dynamics in terms of recruitment of PhDs and Post-docs. We have tagged the contracts with respect to the main axis they are related to, though they may be related to other axes as well. The second chart provides information about the distribution of the contracts with respect to the Institute’s axes: More than 70% of the contracts are related to Axis 1 and Axis 4, which highlights the presence of companies that base their business on core AI technologies, and the important trend of businesses concerned with the smart and secure territory. Financially, Axis 1 and Axis 4 represent 45% and 39%, respectively, of the companies’ investments in partnerships with the Institute.

/ LEVERAGING EFFECT OF THE INSTITUTE ON PARTNERSHIPS In addition to economic development and startup creation support, the other leveraging effect is related to the Institute’s ability to attract new companies to the project, as confirmed by the number of contracts signed and the interest expressed for collaboration. Among the 57 contracts, 27 have been with new companies, that is 47% of the signed contracts. Another leveraging effect is the 19 PhDs and Postdocs recruited under research contracts.

Among the 57 contracts, 27 have been with new companies, that is 47% of the signed contracts.

/ 3IA TECHPOOL A SOFTWARE ENGINERING PLATFORM The creation of joint laboratories with industrial partners had been planned, but discussions with our partners have shown the relevance of setting up a system that would encourage interactions on a larger scale. The Institute has therefore chosen to create in 2021 a pool of engineers “3IA Techpool” hosted by Inria and oriented towards collaboration with industrial partners of all sizes on all subjects of interest to them. It is staffed by 3 engineers. As part of a research contract with a company, the Institute’s researchers can call on the 3IA Techpool to perform software development tasks. This kind of collaboration can last from 2 to 6 months, and

3 engineers to perform software development tasks

it encourages longer-term collaboration and technology transfer. Startups from the Start-it-up program

also benefit from 3IA Techpool support for technology maturation, prototyping, or technology transfer. Some R&D collaborations need to start with an exploration of the scientific state-of-the-art to evaluate the algorithms and technologies from research that could apply to the company’s data and use cases.

Valorization / A DEDICATED UCA INNOVATION VALORIZATION COMMISSION Knowledge transfer is one of the ultimate objectives of research by pushing research and technology results into the market. The Institute benefits from the dedicated “UCA Innovation Valorization Commission”, which brings together the Members, and addresses the issues of valuing multi-partner assets from the Institute, in conjunction with other players in the innovation ecosystem, such as the PACA-Est Incubator, SATT Sud-Est, in charge of managing the patenting of research and technology assets of CNRS and UCA, and the BPI France public bank.

Manage IP assets generated through mutualized structures This support is provided as early as possible, even before the launch of the project, in order to establish a technology transfer plan in a consensual manner between the 3IA partners, and to identify the fund-

Identify the lead member that will handle the technology transfer

Identify the funding necessary to complete this plan

Establish a technology transfer plan

ing necessary to complete this plan. A particular effort is made to standardize the valuation and transfer policies of the Members, in particular in cases where

several partners contribute to the development of the valued technology. It is therefore important to define the valuation scheme as early as possible so that the promoters of the business creation project or industrial partners can have a clear vision of the technology acquisition scheme. Therefore, identifying as early as possible the lead Member that will handle the technology transfer is a key factor in the success of the proposed transfer.

Make the Members of the Institute agree on the legal terms and conditions Typically, this commission allows the different Members of the Institute to agree on the way to manage IP assets generated through mutualized structures such as the 3IA Techpool. As these engineers are financed by different Members of the consortium who are not necessarily the lead Member of a given transfer contract, the commission could make the Members of the Institute agree on the legal terms and conditions, which is the simplest scheme for the contracting partner in mind.

/ STATUS ON VALORIZATION There are three kinds of valorization that have been taken into account: Industrial patents, Software patents and Software licensing. The Institute generated the filing of 13 software patents and 8 industrial patents, of which 6 were jointly filed with partner companies. Among the co-owned patents with companies, 3 were the subject of a transfer (Ekkinox, Dassault and Google). The most prolific axis so far is Axis 1 with 48% of valorization operations, fol-

13 software patents and 8 industrial patents

lowed by Axis 2 with 30%. The Start-it-up program is meant to accelerate knowledge transfer, by making proposals conditional on the transfer of technology.

We expect the numbers of patents or licenses to increase by at least the number of startup projects for the Start-it-up program, which has reached 2 at the end of 2021, with 3 to 5 more planned in 2022. Valorized Item distribution

Valorization per Axis

35%

56%

Licensing

Patent

13%

Software

30%

48%

Axis 1

Axis 2

Axis 3

Axis 4

/ THE START-IT-UP PROGRAM The collaborations and discussions initiated during the first 2 years have revealed the need to support startup creation projects resulting from or related to the Institute’s research. Therefore, the Institute opened in 2021 a call for “Start-it-up 3IA” projects. Since its inception, 6 startup projects have submitted a request for support, 3 of which have been approved and 3 others are in review. Selected projects will receive support for the technological pre-

6 startup projects submitted

maturity phase, and eco-

3 approved 3 under review

entrusted to the PACA-Est

nomic maturation will be Incubator. Six to eight start-

up projects will be supported by 2023.

By the end of 2022, 6 start-ups coming from 3IA Côte d'Azur are expected.

/ DISSEMINATION AND MARKETING ACTIONS The “3IA demonstrators” initiative The Institute launched the “3IA demonstrators” initiative, an online digital showcase providing access to technology demonstrators developed by the Institute’s research. Some demonstrators are intended for the general

RESEARCH

I N N O VAT I O N

E D U C AT I O N

3IA DEMO

A I for REAL WORLD

public and others are targeted to the technically advanced public. The Institute also contributes to dissemination actions initiated by our partners, such as a contribution to the mapping of AI resources (research, training courses) in the region on behalf of the Chamber of Commerce and Industry of the Alpes-Maritimes. The Institute is also the main stakeholder in the Maison de l’Intelligence Artificielle (MIA) in Sophia Antipolis. Established by the Alpes-Maritimes Departmental Council, the MIA is open to the public and is dedicated to helping everyone discover AI, especially students from the Department’s high schools.

3IA Webinars for companies: keep in touch during 3ia.univ-cotedazur.eu the “covid crisis” @3IAcotedazur.fr

3IA cotedazur.fr

The Institute organizes meet-ups with companies through webinars and networking meetings to communicate abouts its offer and generate leads. Due to the health crisis, the institute could have feared a drop in investments from partner companies. While some of them had to postpone their investments in AI R&D, the efforts made to remain in contact with all the partners in the ecosystem allowed us to maintain

6 webinars 150 participants 70% of companies

the overall level of commitment. During the health crisis, relations with the business community were established through webinars: 6 webinars were organized for a total of 150 participants, 70% of which were companies.

How 3IA students serve companies: “Maths Study Group with Industry” The Institute participates in initiatives organized by its Members to gain visibility and attractiveness. For example, the Institute organized in October 2021 a “Maths Study Group

5 companies 5 problems to solve 25 students involved

with Industry”, an event that brings students and companies together for a week. Promising results

were achieved on several subjects submitted by 5 companies.

INTERNATIONAL EXPOSURE AND RELATIONSHIPS

International challenges The Institute first of all contributes to reinforcing the visibility of the Université Côte d´Azur community at the European level, in particular through its participation in the "Ulysseus" European University project, of which AI is a priority theme. An Innovation Hub dedicated to AI aims to federate academic AI skills in the alliance, but also ecosystems more broadly, so as to be able to feed research and innovation in AI at the European Union level, and to position member institutes together for future calls for projects of the European Commission.

AI4media project, part of the H2020 European program The Institute is also involved in the AI4media project, a Center of Excellence founded as part of the H2020 European program, which aims to develop artificial intelligence tools for the media, multimedia and audiovisual industries. As part of this project, our Institute is actively participating in the "International Artificial Intelligence Doctoral Academy" (AIDA), the objective of which is to become a European benchmark in teaching AI at the doctoral level. It should be noted that the first course of this academy was given in the Fall 2020 by Prof. Marco Gori, an international chair of the Institute.

International bilateral collaborations with Universities Internationally, the Institute shines through the bilateral collaborations of chairholders, and through institutional collaborations, in particular the strategic partnerships of Université Côte d´Azur with Université Laval in Canada (co-supervision of 3 doctoral theses in AI, a collaboration which resulted in a publication in Nature Communications, a MITACS Global grant for a UCA student who will visit Université Laval for 3 months, etc.), and with the University of Santiago in Chile (UCA-CNRS-INRAE) with which joint research in modeling is applied to agriculture and biocontrol. Discussions are also underway for the development of Franco-German partnerships with Munich, with the Fondazione Bruno Kessler in Italy, and with the Insight Centre in Dublin.

3IA Côte d'Azur, stakeholder of the Digital Innovation & Intelligence Artificial Lab The Institute and OTESIA are also stakeholders of the Digital Innovation & Intelligence Artificial Lab, created in 2020 as part of the U7+ alliance, and whose initial work led to the drafting by the participating universities of a common position paper on the “Role of universities in the production of ethical and responsible AI”. Finally, we should mention events such as the SophI.A Summit conference and the Deep Learning School, which each year attract around 500 researchers, academic partners and students from different regions of the world.

The Institute thus intends, through various partnerships and visible actions, to strengthen its international positioning. The Institute has also set up several tools to enhance its international appeal:

The Institute has already hired 3 international chairholders They will spend 12 months on the Côte d’Azur to develop research collaborations and provide teaching: •

Prof. Marco Gori (Siena Artificial Intelligence Lab, Italy)

•

Prof. Sebastien Ourselin (King's College London, UK)

•

Prof. David Wales (University of Cambridge, UK)

Three additional international chairs will open in 2022.

The Institute also offers several visiting professor packages each year. There has been poor uptake of this scheme since 2019 due to the health crisis (7 invitations were issued and only 4 of them were accepted due to international travel restrictions). The program will be relaunched in 2022. It is worth noting that all these actions for the visibility of our Institute, and more generally of the network of 3IA Institutes, seem to be effective as the “3IA model” is quite visible abroad.

Several foreign states or research agencies have contacted us to find out about this model of institute, in particular the State of Sao Paulo (Brazil), Saudi Arabia, Israel and Japan. A first visit of the Japan Science and Technology (JST) research agency was held in December 2021.

GOVERNANCE & BUDGET

Governance and budget The 3IA Côte d’Azur Institute is coordinated by Université Côte d’Azur in partnership with the key highereducation, research and innovation institutions in the Nice Sophia Antipolis region: Inria, CNRS, INSERM, and EURECOM. In December 2021, SKEMA Business School, which had supported the project since its inception, formally joined the consortium. MINES ParisTech, which participates in 3IA Côte d’Azur efforts through joint research actions and projects around training in AI, has also expressed its desire to join and formal discussions are in progress. The overall budget of the Institute was €49.5M over 4 years when the project was approved. With SKEMA joining the consortium, the global budget now amounts to €51.5M

Governance

/ EXECUTIVE COMMITTEE The management is ensured by the director assisted by an executive director, a scientific director and a deputy scientific director. This Executive Committee makes all the decisions, after consulting with the restricted steering committee.

/ INSTITUTE COUNCILS The Restricted Steering Committee is a key body: It includes the representatives of the partner institutions and meets every two weeks to discuss the strategic orientations and the actions of the Institute. In an extended form (Extended Steering Committee), it includes all the structures involved in the Institute (such as labs or graduate schools) and other representatives of the ecosystem. This extended committee meets twice a year to give advice on the Institute’s positioning within the overall UCA community policy.

The Scientific Council leds by the Scientific Director. Its mission is to give advice on the scientific and training program, and to referee the selection of new Chairs, PhD students and post-doctoral researchers.

Local Coordination Committee : In order to better coordinate the Institute’s strategy with respect to the ecosystem, a monthly Local Coordination Committee led by the director of 3IA brings together all the public actors of the Region such as local authorities, industrial network representatives, the MIA, and the observatory OTESIA and coordinates AI-related activities in the region.

Industrial Council: For advice on its innovation program, the Institute calls this Council composed of representatives of partner companies and clusters. It is headed jointly by the Institute’s director and by a representative of a partner company.

An Advisory Committee should be set up in 2022. It will be composed of members external to the consortium and will include at least 1/3 of industry representatives. The constitution of this committee will match the strategic committee of the UCA community. This committee is expected to meet once a year to give its opinion on the Institute’s orientations.

/ OPERATIONAL LEVEL At an operational level, the Executive Committee is assisted by those in charge of specific actions: a proj-

ect manager for all the Institute’s activities, two coordinators for relations with companies, a training coordinator, and a digital communication officer. The members of the Executive Committee and the people in charge of actions meet regularly to discuss the strategy implementation. A seminar was also organized in 2021 for all Member staff involved in implementing the project (more than 50 people) at the initiative of the Institute for the purpose of strengthening the sense of belonging, reinforcing coordination around the project and reflecting on actions undertaken together.

An administrative and financial coordination unit headed by the executive director of 3IA shares information with the administration of partner institutions: rules for the use of resources and recruitment; legal documents (consortium agreements, conventions, etc.); organization of the declaration of expenses for reimbursement and financial statements and monitoring of the project progress indicators.

A Business Relations Coordination Unit, headed by the 3IA coordinators, meets regularly to monitor the Institute’s relations and contracts with its partners. A common CRM tool has been set up for this purpose. This Unit also supervises and coordinates the Techpool’s activity. The follow-up of training courses and students is ensured by the Training coordinator of the Institute in collaboration with the partner institutions and the components of Université Côte d’Azur. A

pedagogical engineer has re-

cently been hired to strengthen the 3IA student network.

/ COMMUNICATION A shared unit has been set up to coordinate the initiatives of all partners in order to maximize the impact of communication about the Institute’s actions. This unit is naturally well connected with similar services of the network of 3IA institutes.

/ HARDWARE-SOFTWARE PLATFORMS Quality and ambition of the hardware-software platforms: all researchers of the Institute (chairholders and their teams) have access to the computing infrastructure (OPAL project) shared by Institute members, which already offers a reasonable computing capacity for AI and HPC. In addition to these local infrastructures, access is provided at the national level to France’s most powerful supercomputer for research, named after Jean Zay. The Institute contributed to OPAL, the local computing infrastructure, by purchasing a cluster node with several GPUs, which will be used in priority by 3IA researchers and engineers.

Budget / PARTNERS’ CONTRIBUTIONS TO THE OVERALL BUDGET The global budget now amounts to €51.5M,

PIA grant increase requested after mid-term evaluation

broken down as follows: €3M

•

potential increase in investment by 2023

€19.5M will be contributed by the public partner institutions.

•

€3M

€16M is expected to be funded by the French State through the ANR. Of the €16M

Investment increase achieved in 2021

€10,8M already raised in 2021

€16M

funded by the ANR, €4M is conditional on

€16M

required by 2023

the justification of contributions valued at 16 M€ by the partner companies, €4M of which must be in cash.

Members

PIA Grant

Companies

/ BUDGET STRATEGY Human Ressources expenses HR expenses, the largest budget item, were committed during the first year of the project so that the activity of the Chairs could start quickly. The recruitment of PhD students and post-doctoral researchers as well as the bonuses of the Chairs are financed from the PIA grant. Operating packages, co-financed by the PIA and by the partner institutions, are made available to the Chairs each year.

Contributions of the industrial partners Other resources will be financed by contributions of the industrial partners as defined in the contracts. At the start of the project, 62 companies supported 3IA Côte d'Azur by committing to carry out collaborative research and development projects with the Institute for a total estimated amount of €18.7M. To date, 3IA Côte d'Azur is in contact with more than 140

In contact with +140 companies for actual or planned collaborations. 57 signed contracts + €4.27M in cash to the Members. €10.8M of investments from companies

companies for actual or planned collaborations. 57 collaboration contracts have been signed with companies, bringing in more than €4.27M in cash to the Members. At the end of 2021, the investment of these companies for these contracts is more than €10.8M. The potential investment of companies is estimated at €19M by 2023. In order to further increase the momentum, the Institute plans to set up an incentive system for researchers who engage in collaborations with companies. This system will be financed by the resources provided by the companies.

Other funds The Institute also aims to raise other funds, notably through grants (CASA, for example, provides an annual grant of €30,000 to 3IA), or by responding to national or European calls for proposals for Chairs or structuring projects (e.g., a project submitted in response to the call for expressions of interest "Digital demonstrators").

69 % of the overall budget has been committed to date, which demonstrates progress in line with the initial roadmap. The actions already undertaken, the increase in contributions from academic partners and the results already achieved in terms of business investment seem to argue in favor of a re-evaluation of the PIA grant up to 19 M€. This reassessment - subject to the jury's approval - would allow the Institute to continue to build the momentum already initiated in favor of its ecosystem.

WHAT’S NEXT ?

/ MID-TERM PROJECTION - 2 YEARS Human Resources For the next couple of years, we plan to support the momentum in AI established during the past two years by our Institute, until the end of the probationary period. We will, in particular, renew the startup program

Additional research engineers and an assistant professor

for an additional year, and we plan to hire one or two addi-

Renew the junior Chairs and open new more

the research activity of the Institute, we also would like to

tional research engineers and an assistant professor, for which the budget is secured. In order to keep enhancing evaluate and renew, if appropriate, the junior Chairs whose positions will come to an end in 2023 and open one or two

more new Chairs before the end of 2023. These new Chairs will naturally require additional resources (PhD's, post-docs, bonus, etc.).

Scientific strategy From the scientific point of view, we envision that these new Chairs will pursue the interdisciplinary opportunities of the Institute. For all these additional actions, we estimate the need for complementary funding of €3M from PIA, which

Need for complementary funding from PIA: €3M

is in accordance with the recommendations of the jury after the initial evaluation.

Education Regarding education, a few additional new courses developed during the last two years will be opened by 2023 (introductory course in AI for Humanities, degree in AI & Health with the national Connected Campus initiative, etc.), and will complete our training offering.

A unique place on the Inria campus We also plan, in the coming two years, to find a dedicated and unique place on the Inria campus to welcome our team of engineers, and our scientific visitors. This place could also be used to present the Institute’s research demonstrators and serve as a co-working place for PhD students and researchers from the Institute and associated companies.

/ LONG-TERM PROJECTION - 10 YEARS In the long term, we want to expand both the thematic scope and the size of our Institute to reach the expectations of our local, national, and international partners.

Increase the number of Chairs Firstly, we plan a realistic but significant increase in the number of Chairs of the Institute. This increase should in particular balance the planned retirements of some of our chairholders (11 retirements before 2030) by allowing young and new researchers to benefit from a chair. Chairs may be renewed, based on the scientific committee’s evaluation of a report presenting the achievements of the chair in the first term and the new research questions to be addressed. The number of renewals for the same chair will be limited to two. We therefore plan to bring to an end one or two Chairs each year (e.g., due to retirements, lack of results, etc.), and in the meantime, to recruit between two and four new chairholders per year. Thus, within 10 years, the Institute should have about 60 chair-

Within 10 years, the Institute should have about 60 chairholders Expected budget for this plan: around €6.5M/year

holders for research, education, and innovation. The expected budget for this plan is estimated around €6.5M/year, starting from 2024. Part of this budget will be directly supported by private partner companies. In this regard, a new model of chair is envisioned, co-constructed between a researcher and a company sharing a common AI research project. In accordance with the growth plan for the Chairs, we also aim at widening the scope of the Institute’s research topics by developing a few new themes. In particular, we aim to hire new Chairs on cybersecurity, robotics, digital humanities, and digital ethics, to reinforce interdisciplinarity within the Institute. We also intend to strengthen the core AI axis by welcoming new researchers in key fields such as NLP,

A new model of chair: co-constructed between a researcher and a company sharing a common AI research project New Chairs on cybersecurity, robotics, digital humanities, and digital ethics

Deep Learning, edge AI, or explainable AI.

Enhance the engineering team of the Institute Secondly, we plan to enhance the engineering team of the Institute, since it is a key element in the collaboration with private companies and local authorities, and encourages multidisciplinary collaborations. Having a team of trained AI engineers is an efficient way to increase the collaboration capacity of chairholders while allowing research products to be transferred to large companies or startups.

Create new education programs Finally, in accordance with the massive demand from students, partners, and the new phase of the French national AI plan, our ambition is to create new education programs, based on those developed within the network of 3IA Institutes. Possible actions may include the massification of introductory AI courses using digital platforms, advanced AI courses dedicated to specific professions (Medicine, Biology, Aeronautics,

Microelectronics, etc.), and dedicated programs for large French companies to help their engineers and technicians upskill to the latest advances in AI. It will also require the French AI Institutes to be equipped with digital platforms for teaching AI modeling and algorithms on a large scale, allowing a reduction of the equipment costs.

APPENDIX

Research unit partners of the project Name of the institution

Acronym

INRIA INRIA INRIA MINES Paris-Tech INRIA Eurecom Eurecom Eurecom INRIA CNRS-UCA INRIA CNRS CNRS - INSERM UCA CNRS - UCA Eurecom Eurecom Eurecom OCA - CNRS - UCA UCA - CNRS UCA - CNRS INRIA CNRS INRIA INRIA INRIA UCA INRIA INRIA SKEMA

ABS ACUMES Athena CMA Datashape Data Science department Digital Security department Communication Systems department Epione GREDEG HEPHAISTOS I3S iBV IPMC Data Science department Digital Security department Communication Systems department LAGRANGE LEAT LJAD MAASAI MDSC Morpheme NEO Stars TIRO-MATOs Titane Wimmics SKEMA AI Institute

PhD Students JULIEN AUBERT (CHAIR OF P. REYNAUD-BOURET) CNRS, LJAD Estimating the learning process of the brain ALI BALLOUT (CHAIR OF A. TETTAMANZI) UCA, iS3S Active Learning for Axiom Discovery CHRISTOS BOUNTZOUKLIS (CHAIR OF E. DI BERNARDINO) UCA, LJAD Forest fire risk mapping using geospatial analysis and machine learning JULIE CHARPENET (CHAIR OF M. TELLER) UCA, GREDEG Freedom of speech challenged by the moderating power of digital platforms ANTOINE COLLIN (CHAIR OF P. BARBRY) IMPC, CNRS Automatic celltype annotation for atlas construction RYAN COTSAKIS (CHAIR OF E. DI BERNARDINO) UCA, LJAD Open problems in spatio-temporal extreme-value modeling HIND DADOUN (CHAIR OF N. AYACHE) Inria, Epione AI-based real-time diagnostic aid for abdominal organs in ultrasound ANTONIA ETTORRE (CHAIR OF CATHERINE FARON) UCA, I3S Artificial Intelligence for Education and Training: Knowledge Representation and Reasoning for the Development of Intelligent Services in Pedagogical Web Environments ZHIJIE FANG (CHAIR OF N. AYACHE) Inria, Epione Statistical Learning for multimodal registration medical images ETRIT HAXHOLLI (CHAIR OF M. LORENZI) Inria, Epione Exploring latent dynamical models for failure prediction in time-series of high-dimensional and heterogeneous data ASHWIN JAMES (CHAIR OF ALEXANDRE MUZY) CNRS, I3S Modeling and neuromorphic simulation of learning individuals based on the interaction between behavioral and neuronal activities

VICTOR JUNG (CHAIR OF J-C. RÉGIN) UCA, I3S Learning of Hidden Constraints VICTORIYA KASHTANOVA (CHAIR OF M. SERMESANT) Inria, Epione Learning Cardiac 3D Electromechanical Dynamics with PDE-based Physiological Constraints for Data-Driven Personalized Predictions in Cardiology YACINE KHACEF (CHAIR OF C. RICHARD) UCA, Lagrange Distributed dark fiber optic sensing for smart cities monitoring BOGDAN KOZYRSKIY (CHAIR OF M. FILIPPONE) Eurecom Freedom of speech challenged by the moderating power of digital platforms HUIYU LI (CHAIR OF H. DELINGETTE) Inria, Epione Machine Learning Methods for the Anonymization of health Data DINGGE LIANG (CHAIR OF C. BOUVEYRON) UCA, Inria - MAASAI Generative deep learning for network analysis DAVID LOISEAUX (CHAIR OF J-D. BOISSONNAT) Inria, Data Shape Multivariate topological data analysis for statistical machine learning STEVE MALALEL (CHAIR OF J-C. RÉGIN) UCA, I3S Decision-making process in a multi-objective environment including stochastic data GIULIA MARCHELLO (CHAIR OF C. BOUVEYRON) UCA, Inria, MAASAI Statistical learning from dynamic bipartite networks with heterogeneous edges SANTIAGO MARRO (CHAIR OF S. VILLATA) CNRS Argument-based explanatory dialogues for medicine ARTEM MULUKIOV (CHAIR OF B. MIRAMOND) UCA, LEAT Confrontation of self-organizing maps and spiking neural networks for embodied computation inspired by the brain ANGELO RODIO (CHAIR OF G. NEGLIA) Inria, NEO Sustainable distributed machine learning HUGO SCHMUTZ (CHAIR OF O. HUMBERT) UCA, Centre Antoine-Lacassagne Predicting immunotherapy response of lung cancer patients using deep semi-supervised learning

BORIS SHMINKE (CHAIR OF C. SIMPSON) CNRS, LJAD AI for finite algebraic structures VASILIKI STERGIOPOULOU (CHAIR OF L. BLANC-FÉRAUD) CNRS, I3S Learning and optimization for 3D+T super-resolution in fluorescent microscopy VALERIYA STRIZHKOVA (CHAIR OF F. BREMOND) Inria, STARS Emotion Detection using Deep Learning AMIR HOSSEIN TAKAVOLI (CHAIR OF S. DEMASSEY) MINES, CMA Hybrid combinatorial optimization and machine learning algorithms for energy-efficient water networks ROMAIN TISSOT (CHAIR OF J-P. MERLET) Inria, HEPAHISTOS Using IA and cable-driven parallel robots for the assistance to frail people and medical monitoring PAUL TOURNIAIRE (CHAIR OF H. DELINGETTE) Inria, Asclepios AI-based selection of imaging and biological markers predictive of therapy response in lung cancer ATHANASIOS VASILEIADIS (CHAIR OF F. DELARUE) UCA, LJAD Mean field approach to MARL (Multi Agent Reinforcement Learning) CEDRIC VINCENT-CUAZ (CHAIR OF R. FLAMARY) UCA Optimal Transport for structured data ALEXANDRA WÜRTH (CHAIR OF P. GOATIN) Inria, Acumes AI for road traffic modeling and management TONG ZHAO (CHAIR OF P. ALLIEZ) Inria, Titane Learning priors and metrics for 3D reconstruction of large-scale scenes LIU ZIMMING (CHAIR OF PHILIPPE MARTINET) Inria, CHORAL Representation of the environment in autonomous driving applications RÉMI FELIN (CHAIR OF A. TETTAMANZI) UCA, I3S Evolutionary Axiom Discovery from Knowledge Graph RICCARDO TAIELLO (CHAIR OF M. LORENZI) Inria, Epione Data Structuration and Security in Large Scale Collaborative Healthcare Data Analysis

OUALID ZARI (CHAIR OF M. ÖNEN) EURECOM Adversarial Learning and Differantial Privacy PIERPAOLO GOFFREDO (CHAIR OF S. VILLATA) CNRS, I3S Natural language counter argumentation to fight online disinformation NINA SINGLAN (JEAN MARTINET) UCA Neuromorphic Visual Odometry for Intelligent Vehicles with a Bio-inspired Vision Sensor KEVIN MOTTIN (CHAIR OF M. GORI) UCA, I3S, INRIA - MAASAI Learning to Explain - Explain to Learn (L2EE2L) BENJAMIN OCAMPO (CHAIR OF ELENA CABRIO ) UCA, I3S Automatic detection of abusive content ZAKARYA EL KHIYATI (CHAIR OF JÉRÉMIE BEC) Inria Reinforcement learning for the optimal locomotion of micro-swimmers in a complex chaotic environment FRANCESCO GALATI (CHAIR OF M.A ZULUAGA) EURECOM Interactive and Collaborative Learning for 3D Vessel Segmentation STEFANO SPAZIANI (CHAIR OF P. REYNAUD-BOURET) CNRS Feasibility Engagement for NeuroMod application

Post-doctoral researchers ALESSANDRO BETTI (CHAIR OF M. GORI) UCA, Maasai Lifelong Learning in Perception SUHANYA JAYAPRAKASAM (CHAIR OF D. GESBERT) EURECOM Learning and coordination at the wireless edge in future Iot EDOUARD BALZIN (CHAIR OF C. SIMPSON) UCA, LJAD Interactions between Pure Mathematics and Artificial Intelligence ANDREA CASTAGNETTI (CHAIR OF B. MIRAMOND) UCA, LEAT Deep spiking neural networks for image denoising in smartphones WILLIAM HAMMERSLEY (CHAIR OF F. DELARUE) UCA, LJAD Stochastic Partial Differential Equations, Mean Field Stochastic Particle Systems, Theoretical Analysis of Machine Learning Algorithms KRISTOF HUSZAR (CHAIR OF J-D. BOISSONNAT) Inria, Data Shape Computational Geometry Learning in Non-Linear Spaces DANIEL INZUNZA (CHAIR OF P. GOATIN) Inria, ACUMES Advanced data-driven modelling for traffic flow accounting for emerging technologies THI KHUYEN LE (CHAIR OF O. HUMBERT) UCA, TIRO Deep learning for the diagnosis of brain tumor recurrence based on 18F-DOPA PET images VLADIMIR KRAJNAK (CHAIR OF F. CAZALS) Inria Decoding the energy landscape: understanding and redesigning the emergent properties of a kinetic transition network STEFAN SARKADI (CHAIR OF F. GANDON) Inria, Wimmics Knowledge-sharing in Open Multi-Agent Systems AUDE SPORTISSE (CHAIR OF P-A. MATTEI) Inria, Maasai Handling missing data with deep learning, applications to cancer prognosis

MOHSEN TABEJAMAAT (CHAIR OF F. BREMOND) Inria, STARS View invariant action recognition JOSUÉ TCHOUANTI (CHAIR OF P. REYNAUD-BOUYRET) CNRS, LJAD Testing Mean-Field behavior AYSE UNSAL (CHAIR OF M. ÖNEN) Eurecom Adversarial learning and anonymisation for federated machine learning MARTIJN VAN DEN ENDE (CHAIR OF C. RICHARD) UCA, LJAD & Géoazur Distributed dark fiber optic sensing for smart cities monitoring SIXIN ZHANG (CHAIR OF L. BLANC-FÉRAUD) CNRS, I3S Deep Learning for extracellular matrix analysis from biological multispectral images MAURO ZUCHELLI (CHAIR OF R. DERICHE) Inria, Athena AI-based Computational Brain Connectomics

3iA Activity Report 19-21

Articles inside

GOVERNANCE AND BUDGET

INTERNATIONAL VISIBILITY

APPENDIX • Research unit partners of the project • List of 3IA PhD's and Post-docs

WHAT’S NEXT ?

AI for smart and secure territories

The 3IA national network

Editorial

What’s 3IA Côte d’Azur ?

AI for integrative computational medicine

Local dynamic

AI for computational biology and bio-inspired AI

RESEARCH

Scientific publications