The scientific challenges dealt with in the research programs, resulting from the chair projects and use cases provided by ANITI's industrial partners, are broken down into twelve themes.
Certain themes are dealt with jointly within several research programs.
Learning with few or complex datas
This theme explores new learning approaches and architectures inspired by "natural" or "biological" learning mechanisms: Humans and other animals can learn complex information from sparse and often noisy data. Another avenue is the extension of automatic data generation techniques for learning, so that they can be scaled up and applied to problems other than those on which they have been tested so far. Another challenge concerns the definition of compact methods for the representation of large-dimensional data, suitable for deep learning. The ambition is also to develop a unified framework to harmonize multi-source, multi-modal, multiscale and time-dynamic data into a coherent semantic representation for deep learning.
The ambition is also to develop a unified framework to harmonize multi-source, multi-modal, multiscale and time-dynamic data into a coherent semantic representation for deep learning. We will for example apply approximate Bayesian computation to problems of algorithms for extracting hidden properties (described in latent space) in such data. Applications are envisaged in the fields of remote sensing, transport and biology.
Thread 1: Low resource learning
Thread 2: Representation learning
Thread 3: Multi-source, Multi-scale data
Safe design and embeddability
This theme addresses software/hardware architecture, verification and system-level assurance of certifiable AI systems. It draws on results from the themes of learning, optimisation, as well as AI and physical models.
The theme also explores different means of verifying integrated learning models in critical systems in order to guarantee a high level of safety for the system.
One of the objectives is also to contribute to the definition of certification standards for the development of critical systems using AI algorithms.
Thread 1: Embeddable AI architecture
Thread 2: AI Model selection and verification
Axis 3 : Assurance des Systèmes d’IA
Chair holders : Serge Gratton, Joao Silva Marques, Claire Pagetti
Datas & anomaly
This theme aims at defining advanced techniques based on hybrid AI for the detection, diagnosis and prognosis of anomalies or extreme events.
The first objective is to understand how model and data driven approaches can complement each other effectively. The second objective is to be able to abstract data classifiers and match them with symbolic or analytical models adapted to reasoning for diagnosis purpose, in order to obtain a better explainability and acceptability of the results.
Different use cases will be considered to validate the proposed techniques, for example, predictive maintenance and diagnosis of electronic boards, cobots or industrial systems represented by digital twins, flood detection by digital simulation augmented by machine learning techniques.
Thread 1: Semantics from noisy linguistic data
Thread 2: Language and multi-modality Thread 3: Dialog/conversation modeling
Chair holders : Jean-Michel Loubès, Serge Gratton, Jean-Bernard Lasserre, Louise Travé-Massuyès
This topic studies three approaches to improving the explainability of learning algorithms. The first uses a representation of a learning algorithm by a set of first-order logic clauses, from which various notions of explanation can be studied in a general logic framework. The main challenge is to scale these methods.
The second class uses statistical techniques that consist in disrupting inputs and measuring the effect on classifications, thus indicating key learning parameters. This method is efficient but does not formally guarantee the results.
This method is efficient but does not formally guarantee the results. The third class consists in building hybrid combinations of the two methods in order to improve the
quality of the explanation and the performance of obtaining this explanation
Thread 1: Explainability with logical methods
Thread 2: Explainability with statistical methods
Thread 3: Hybrid XAI : combining statistical and logical methods
Chair holders : Leila Amgoud, Joao Silva Marques, Jean-Michel Loubes, Thomas Schiex, Louise Travé-Massuyès
The objective is to develop new methods to detect and then eliminate undesirable biases in learning, validation and test datasets or in probability distributions associated with a learning architecture, which the end user can specify.
This topic also examines how these methods meet the legal and ethical requirements of AI systems.
Of particular interest is the application of these methods to critical industrial applications for which bias may arise from an unbalanced representation of operating and environmental conditions, mislabelling or incomplete description of data leading to erroneous correlations.
Thread 1: Analysis of bias for fairness
Thread 2: Analysis of Bias for Data and algorithms in critical system design certification
Chair holders : Leila Amgoud, Celine Castets-Renard, Fabrice Gamboa, Jean-Michel Loubès, Claire Pagetti, Joao-Marques Silva.
AI and physical models
Analytical models for analyzing complex processes (physical, chemical, biological, …) like the planet’s weather, known as physical models, are often incomplete or intractable. Thus complementing and even replacing such models with data-driven ones using machine learning has become a popular approach in many fields of science and engineering. ML techniques have been shown to be able to solve complex classification or regression tasks very efficiently for data such as images, text, audio signals. This theme investigates how ML techniques can be extended to solve problems, involving Physics represented by raw data or by partial differential equations. It features a hybrid approach to AI mixing physical constraints with machine learning methods, energy models, etc.
The objective is to accelerate the simulation of these models or to improve the learning of deep models under physical constraints, by combining approaches from pure mathematics (algebra, topology, geometry), applied mathematics (statistics, numerical methods) and data
Thread 1: Accelerating physical models simulation and optimization with ML, using statistical and geometric, approaches
Thread 2: Improving learning models with physical constraints
Chair holders : Serge Gratton, Fabrice Gamboa, Thomas Schiex, Jérôme Bolte, Nicolas Dobigeon
AI & society
This theme addresses the challenges related to the socio-economic, legal or ethical acceptability of AI applications, answering questions such as "how does AI affect economic competitiveness, economic platforms? "How can we use AI to better involve the general public in social/governmental decisions? How are autonomous AI systems perceived by the general public, how can we assess the associated risks and how can we make them more acceptable?
The work is also related to legal concerns and responsibilities and privacy protection.
Thread 1: Social acceptability and applications of AI
Thread 2: Responsibility: Legal and Ethical Issues
Chair holders : Jean-François Bonnefon, Bruno Jullien, Cesar A. Hidalgo, Céline Castets-Renard
Contributions focus on the extraction of deep semantic information from texts in semi-closed domains (e.g. maintenance or production logs) or from noisy conversations.
We exploit not only lexical semantics, but also semantics that is encoded in semantic relations between clauses or discourse structure. The latter is important for capturing, for instance, reasons why an operation was or was not performed, exceptions to requirements, elaborations or more detailed descriptions on proposals or operations performed, but also opinions and comments.
We also study the foundations of grounding and multimodal representation learning in conversation (language, vision). These advances in dialogue and conversation modeling will improve the performance of robots/cobots and virtual assistants that have access to visual and linguistic data, and will also be exploited to produce sophisticated conversational systems that are embedded in a physical environment to learn and execute new complex actions in interaction with humans.
Thread 1: Semantics from noisy linguistic data
Thread 2: Language and multi-modality
Thread 3: Dialog/conversation modeling
Chair holders : Rachid Alami, Leila Amgoud, Thomas Serre, Rufin VanRullen
Nerusocience & AI
This theme aims at cross-fertilization between neuroscience and artificial intelligence techniques. Two main objectives are targeted: 1) to use machine learning methods to monitor in risky environments relevant mental states of users interacting with each other and with artificial agents (workload, fatigue, stress, risk of error); 2) to use knowledge about the brain's modes of functioning to improve artificial intelligence algorithms and neural networks architectures, in particular for reinforcement learning.
Axis 1 : IA pour monitorer et comprendre l’activité cognitive
Axis 2 Reverse engineer the brain to improve AI algorithms
Chair holders : Rachid Alami, Frédéric Dehais, Nicolas Mansard, Thomas Serre, Rufin VanRullen
Optimization & game theory for AI
This theme aims to study the theoretical foundations of different optimization and game theoretic techniques in order to improve machine learning methods based on neural networks and to better understand, from a theoretical point of view, their conditions of convergence, stability and robustness, as well as the ability to generalize their results. This topic is also central for the certification of AI systems used in critical systems.
Thread 1: Optimization theory for AI
Thread 2: Robustness
Thread 3: Game theory and AI
Chair holders : Jerôme Bolte, Serge Gratton, Jean-Michel Loubes, Jean-Bernard Lasserre, Jérôme Renault, Marc Teboulle
Automated reasoning and decision making
Algorithms that can simultaneously handle numerical optimization problems and logical properties (or constraints) with guarantees or certificates are therefore of primary importance for AI, allowing to impose constraints on the output of learned models. The main challenge is that automated logical reasoning is computationally hard (NP-complete decision) and cannot be approximated. This theme aims at exploring different complementary techniques to reduce complexity and improve scalability (e.g., guaranteed solution of decision/optimization problems on weighted discrete graphical models with constraints (SAT/CSP/NP-complete type), knowledge compilation techniques and algorithms for online optimization of problems dealing with preferences and/or uncertainties, ...).
Applications include for example the optimization of the design of molecules and proteins in the field of health or sustainable development, as well as problems of temporal planning (scheduling) and configuration optimization in industry 4.0.
Thread 1: Algorithms and complexity
Thread 2: Applications to complex systems
Chair holders : Leila Amgoud, Daniel Delahaye, Hélène Fargier, Joao Silva Marques, Thomas Schiex, Louise Travé-Massuyès
Robotic & AI
The work focuses on advanced robots with locomotion capabilities, with legs and/or arms, having to perform tasks
such as walking, grabbing an object, opening a door, etc., and collaborating with humans to act together, share space, and jointly perform tasks.
In addition to challenges concerning the development of autonomous functional
and decision-making capabilities needed to perform tasks in the physical environment and in the presence of uncertainties, topics cover motion analysis and generation, motion and task planning, and human-robot collaboration using hybrid AI methods
This theme also includes research on aspects related to the architecture and verification of autonomous robots as well as their societal impact.
Thread 1: Motion planning and control
Thread 2: Cognitive abilities and communication
Thread 3: Architecture, decision and interaction
Thread 4 Social and Societal aspects of Human-Robot interactions
Chair holders : Rachid Alami, Céline Castets-Renard, Frédéric Dehais, Nicolas Mansard, Claire Pagetti, Thomas Serre, Rufin VanRullen