Why are we interested in artificial intelligence?
Artificial intelligence is becoming increasingly important in our modern world, in more and more common situations: just to mention computer vision, we can talk about facial recognition to unlock your smartphone, voice recognition, automated car driving, anonymized crowd analysis or medical diagnostic assistance, etc.
We can also talk about automatic optimization of the behavior of computer and electronic systems, by adapting in real time to changing operating conditions, such as the embedded sensors used for predictive maintenance. These intelligent sensors learn the usual operating modes of machines from various information (vibrations, consumption, etc.) and then succeed in detecting and predicting failures.
Finally, there are also systems that acquire by themselves an understanding of a more or less complex situation, like the calculators that have learned to play Go and beat the best human players, a performance unthinkable with more classical problem solving techniques.
Today, we have two treatment options for all these situations:
- The first is "cloud computing" (remote data processing), in which all data is sent to servers with high processing capacity;
- the second is "edge-computing" (local data processing), whose importance is both growing and already remarkable. This type of processing offers a number of advantages, such as confidentiality of sensitive data, shorter latency times and greater autonomy than cloud computing, even if resources remain limited.
Keeping abreast of these techniques is essential, both to be able to offer pedagogical content, but also to be able to assess whether these innovative solutions can find areas of application in our industrial achievements.
What does this research involve?
Even if "artificial intelligence" is a fashionable term used to designate many rather vague things (a little story goes: "Automation is what we know how to make machines do, artificial intelligence is what we'd like to make machines do"), we're sticking to a definition that we feel is appropriate to a number of situations: the ability of systems created by human beings to extract from their environment a sufficient quantity of knowledge to successfully carry out given tasks.
Performing these specific tasks can range from answering a simple question to classifying data in a non-trivial way or determining a set of parameters that will optimize the behavior of an automaton or a complex system.
This can be applied to computer vision, signal processing in a broader sense, automated value control, real-time control, etc.
The environment can be of a controlled and formalized type (which makes it possible to establish a link between what we want to obtain as a result for a given input) or it can be outside of any prior representation, calculation or acquisition capacity. This forces the artificial system to find a model that will allow it to optimize its operation (improved response, reduced energy consumption, etc.).
This activity gathers at ISEN Yncréa Méditerranée teachers-researchers with different experiences and sensitivities, because the means to reach the ends described above can be very variable, and require competences in linked and intertwined domains but requiring sharp and heterogeneous expertises (in software and hardware, sometimes both at the same time)
Thanks to this diversity of knowledge, it is possible to implement a complete chain of artificial intelligence (acquisition, processing, analysis and actuation) on a system as well as to embed it on a device with limited resources and especially low energy consumption.
There are many issues related to embedded artificial intelligence. Indeed, artificial intelligence is interesting when computing resources (memory and computing power) are somehow unlimited. However, as soon as we move to the embedded world, it is necessary to take into account several constraints, in particular memory, computing power and especially energy consumption.
This is why the complementary domains which are ours at ISEN Yncréa Méditerranée (microelectronics, contactless communication, network, software, etc.) allow us to consider several solutions for each problem posed, in order to find the one or ones which are the most suitable.
How is R&D carried out?
This R&D is carried out through national and international projects by our teacher-researchers. Students can be involved in related parts of these projects in order to meet a specific educational objective.
Different solutions are evaluated, using multiple algorithms that can be described as "machine learning". These solutions then act as guidelines for the realization of proofs of concept, which are then tested on representative samples (data acquired in-house, or supplied by project partners, in the form of signals, data sequences, images, films, etc.).
The results contribute both to the overall experience pool, as well as to recommending viable solutions to various problems.
