AI

Themes

Program Facilitator : BUCHE Cédric

 

Theoretical

Brain-inspired AI

AI techniques are processes that consume a lot of computing resources. The high solicitation of computing units and memory implies a significant energy consumption, which is a determining factor as to the possibility of using these techniques on self-powered embedded devices (autonomous mobile robots to analyze in real time the data from their own sensors for a reaction for example). The optimization and adaptation of generic IA solutions to the very specific characteristics of such a platform and its application expectations are necessary for their effective implementation under such conditions.

 

Robotic and AI

The combination of artificial intelligence and robotics has given rise to "smarter" and more efficient robots. The algorithms must be embedded, respond in real time and using partial sensor data and bad qualities. At lab-STICC we work on the detection of objects; make learning robots by observation or imitation; optimize algorithms for faster processing. Our applications focus on the use of robots to help humans in their daily tasks. Thus we focus on continuous life-long leraning of robots in stochastic environments and in interaction with humans. We study how robots can build increasingly complex skills using robot based on active learning, artificial curiosity and interactive learning. We are also involved in robocup @HOME (PEPPER robots) and soccer (NAO robots).

 

Interactive Machine Learning

Interactive machine learning merges machine learning and human-machine interaction. While traditional machine learning systems process the data given to them in advance, this research considers that the learning process can benefit from interactions with the environment and with a human, and that the outputs of the system convey significant information. Indeed, humans can contribute to a learning algorithm, including entries in the form of labels, demonstrations, tips, awards or rankings. The interaction is all the more useful as the human can the learning process while adapting its guidance to outputs of the algorithm.

 

Explicability and AI

Within the framework of the explicability, the activities of the lab-STICC relate mainly to the narrative representation. We develop models supported by logical approaches, guided by the work of narratology, psychology of story comprehension, or pragmatics, which aims to describe and explain dynamic scenarios. The systems illustrating these models can be purely generative (narrative generation) or interactive (Interactive Storytelling, dialogue generation).

 

AI for complex decision-making

In order to enable a human to make informed decisions, a downstream phase of processing and extracting knowledge from data is necessary. It involves modeling human preferences and applying methods and algorithms that combine human knowledge with those extracted from the data (data mining) to provide relevant and robust decision support. At lab-STICC, this decision aid is translated into multi-objective optimization, multi-criteria decision support, preference modeling, preference learning, integration of multiple decision makers' preferences into algorithms , or even the consideration of human expertise in search algorithms, machine learning, integration and information quality.

 

Cognitive Sciences and AI

The cognitive sciences are designed to study the cognitive processes of living beings. This goes well beyond neuroscience and incorporates knowledge in psychology, biology or neuroscience as well as social, cultural and linguistic processes. Artificial intelligence is then seen as a tool for establishing models and simulations to support a theory or concept. In return, these theories and concepts can fuel the models of artificial intelligence. In the lab-STICC, the following topics are discussed: the theory of human decision-making simulation, interactive and constructivist learning, life-long learning, intrinsic motivation, connectionist models of memory as well as the paradigm of enaction.

 

 

Applicative

 

AI for the (under)water environment

AI works for the marine environment concern both the sea surface, the water column and the sedimentary interface. The application frameworks are varied and include localization, pilot autonomy, mission monitoring of mobile robots as well as the multi-physical perception of the marine environment, its fauna and flora. Academic work focuses on the development of algorithms as well as the embedding of algorithms in a context of autonomy and complex, fluctuating and uncertain environment.

 

Education and AI

In the field of education, work at the lab-STICC addresses different approaches. First of all, knowledge engineering and semantic web allows: a) To adapt and / or personalize learning activities and interactive dashboards dedicated to learners and / or tutors depending on the context; b) Support collaborative learning that promotes learner autonomy and confidence in lifelong learning by respecting ethical rules and privacy. Then, the automatic learning (Learning Analytics) makes it possible to analyze the traces of interaction of the learners, the tutors, etc. to adapt the learning environment, to give feedback to learners and / or guardians, to evaluate the course and modify it accordingly, etc. For example, interaction traces are analyzed in MOOCs to predict and explain the "dropout" phenomenon to alert tutors, identify difficult parts of the course, and so on. Predictive classifiers for automatic feedback and "explanatory" classifiers for personalized feedback to teachers and learners.

 

Virtual Reality and AI

A virtual reality system allows a user to be immersed, both sensorimotorally and cognitively, in a world in which he can act spontaneously, even naturally, without feeling that the interaction is mediated by a technical device and having the illusion that the world presented to him reacts to his actions. In this context, the application of the AI ​​can be envisaged when one wishes to confer on the system a form of adaptability to the user or that the virtual environments must be populated with entities "intelligent", c that is, able to act in environments in a form not specifically scripted for this or that configuration of the environment. The first use case is not entirely specific. It is for the system to be able to recognize the actions of the user, and possibly to build a model, so as to propose an adaptation of the device to the user (for example to facilitate interaction ). The problem is all the more complicated as the user is equipped with few sensors, which is the current trend in virtual reality or augmented. The applicable AI techniques are those of artificial learning. The second use case is that of the design of digital entities to which users will lend a form of intelligence. These are primarily autonomous virtual humans capable of interacting with users; they may also be animals with relatively complex behavior, or even imaginary entities (applications in the field of the arts). Currently, the majority of virtual environments are static (only the user acts) or are populated with entities whose behavior control program is fully scripted, or reactive (synchronized automata events). Interactivity between the user is then low. The ambition we pursue at the lab-STICC is to provide these entities with a form of learning that gives them robust behaviors; moreover, we do not want to have to explicitly program these behaviors. In this framework, we work on behavioral architectures based on cognitive science models ("perception-actions" paradigm and ecological approach to cognition). These models are sources of inspiration for the control of motor behaviors (for example the coordinated movements of small groups of pedestrians), for the coordination between users and autonomous entities (for example, the management of the turns of words between virtual humans and " using non-verbal cues) or for behavioral adaptation to promote user engagement in the interaction (eg for interaction with virtual animals). In this work, we focus more specifically on model-based approaches.

 

Health and AI

In the field of health, the Artificial Intelligence work at the lab-STICC focuses on the use of Clinical Decision Support Systems (CDS) whose purpose is to guide the clinician expert in the development of diagnosis and in the therapeutic orientation. The issues are twofold, improving the quality of care while reducing treatment costs.

 

AI for Ocean Monitoring and Surveillance

This research theme aims to develop novel AI-based techniques for ocean monitoring and surveillance. From a methodological point of view, the focus is given to the development of data-driven and learning-based statistical representations and models for the characterization, reconstruction and simulation of complex dynamical systems governed by ODEs or PDEs (Ordinary and Partial Differential Equations). Applications to ocean monitoring and surveillance include among others geophysical dynamics, ocean remote sensing, movement ecology as well as maritime traffic surveillance.
©2016-2019 Azimut - Website design & Interactive kiosks Legal information | Site map