"To each his own definition of artificial intelligence!"
While AI is increasingly present in our lives and in our imaginations, with its procession of legitimate questions and fantasies, its technological reality remains opaque for many of us. Neural networks, deep learning, algorithms... A brief review of concepts with Anne Laurent, Director of theMontpellier Institute of Data Science and Vice President for Open Science and Research Data at the UM.
Technically speaking, how would you describe how an AI works?
The generic concept of AI refers to the ability of a machine to reproduce our cognitive abilities (reasoning, learning, recognizing, creating...). But the focus is often on learning. The idea is to make a machine learn a notion by training it to distinguish situations from examples, just as humans do. There are many different learning methods. The input training data, the examples, can be photos of moles, for example. The system must then distinguish between melanoma and non-melanoma in its output.
But what happens between this entry and exit?
Today's most powerful learning methods are based on neural networks with highly complex systems and processing. The basic concepts are similar to the way our neural system works. The input signal activates one neuron before being passed on to the next, on several levels, known as layers of neurons. Between input and output, the system will decide whether or not to transmit the signal to the next neuron, and when to do so, in order to accomplish its task. This is what we call neural network learning or deep learning. Deep means deep, because there are many layers of neurons and connections that need to be parameterized. So, for the machine to perform its task successfully, it needs to be trained with huge amounts of data.
Where does this data come from?
When we talk about AI in research, we're talking about research data from, for example, scientific imaging instruments, DNA sequencing... This data is produced either by the researcher using the AI, or by other teams who have worked on the subject and are willing to share it for learning purposes. That's the importance of open science.
Many mathematicians and computer scientists are working on AI by creating new algorithms. What is an algorithm and what is its role?
When you want to make a machine do something, you have to talk to it in its own language, a programming language: Python, Scala or Java, for example. But you don't code or program without thinking. You start by writing an algorithm, i.e. a formalized description, in a language that can be understood by humans, of the data to be used, the operations to be performed by the machine and the conditions under which it will be made to do so.
If we had to define artificial intelligence, what might that definition be?
The term AI was first coined in 1956, and since then it has had its own history and evolution, with some sub-themes developing more than others. In my opinion, to define AI, we need to go back to the initial vision of its founding fathers, who attributed to it the ability to reproduce our cognitive skills: planning, reasoning, deciding, learning, sensing the world, etc. But we also need to take into account the breakthrough brought about recently by computerized systems. But we must also take into account the breakthrough recently brought about by commercial generative AI systems. So, for many people, AI equals generative AI; for others, AI equals statistics... To each his own definition of AI!
For the past few months, AI news has been saturated with Chat GPT, a generative AI. What is generative AI?
This is a form of AI that can generate content: text, images, video... It relies on the input it is given, to which it adds everything it has managed to learn in its model. The user can guide it by giving it context and purpose. For example, he might ask it to give him the causes of melanoma, but addressing teenagers who are going to spend their summer on the beach, or their parents so that they can better educate teenagers. Generative AI contextualizes and adapts the discourse to the required purpose. AI fascinates us because it speaks to us all.
What other types of AI are present in our daily lives?
Some AIs are dedicated to so-called "classification" tasks. I've already mentioned moles, but I could also have mentioned Pl@ntNet, a plant recognition application from Montpellier that can classify plants into the thousands of categories that exist in plant taxonomy. Then there are the so-called segmenting AIs, widely used in marketing to categorize customers without any preconceived ideas. There are AIs that plan and will, for example, help organize schedules by solving highly complex combinatorial problems and constraints. Others make recommendations, on online video platforms for example... There are a huge number of sub-categories, and all these tasks will become increasingly hidden. AI is using more and more models, and sometimes we don't even realize it...
One of the central questions in the development of AI is that of control. Can we verify and explain the work of AI?
It's a highly topical issue, because we have to do things responsibly, and this is now imposed on us by the European regulation passed this summer. Having said that, we need to differentiate between different levels of control. For example, the stakes and methods will differ depending on whether we want the end user to be able to understand the AI's decision, or whether this explicability is more a matter for experts, for example in a court of law. There's always a little tune in my head that tells me not to ask more of AI than we ask of humans. You can't expect it not to make mistakes, not because it's imperfect, but because life isn't binary.
Biases in artificial intelligence results are often mentioned. Can they be limited?
This is usually due to representation bias in the input data. If, for example, we want to teach an AI to recommend study sections to high school students, and we simply reproduce the existing statistics, the AI will never or very rarely recommend the math section to girls. To correct these biases, we need to increase the data to give the machine data where there are as many girls as boys. This is also why France wants to create its own AI models, so as not to be subject to cultural biases that are not our own.
So we can't deliver raw databases to the machine, we have to work on them?
Yes, and that's just the tip of the iceberg, the work that we don't talk about as much, which is very tedious and meticulous, but it makes the data as usable as possible. This is one of the missions of ISDM, which I manage.
What exactly is ISDM's role?
ISDM, the Montpellier Institute for Data Science, supports researchers in the management and processing of research data. It provides tools, infrastructure, training and, thanks to the data clinic, advice and expertise. We also organize the Halles de l'AI, a University initiative to bring together players in the field of AI and data, and we do a lot of work on data storage and security. To sum up, we're a gateway for all those who want to jump on the AI bandwagon, and we need to...
You're also vice-president in charge of open science. What's at stake with the development of artificial intelligence?
Open science is the fuel of this learning process, but sharing does not mean open bar ! Open to all is one way of sharing, but it's not the only way. We can set conditions and restrictions, and think about securing, protecting and promoting our scientific heritage.
Is AI essential for science today?
Yes, absolutely, AI is going to transform the profession of researcher by increasing our capacity to produce a state of the art, to verify, structure or challenge our ideas, and so on. We'll also be able to accelerate the pre-processing of our data, the management of certain administrative tasks or our search for funding...
How is the UM positioning itself in the face of these developments?
Everywhere! The UM has ambitious projects. The aim is to give researchers the ability to take control of all forms of AI, and to show the firepower that AI has to acculturate and change our practices as we go along. All within the most ethical framework possible, without "leaking" our ideas or destroying the planet. The UM also strongly supports AI research and the development of new algorithms. Montpellier is developing leading-edge methods. This is the case, for example, with AI methods in healthcare, such as federated AI, in conjunction withInria. There isn't a huge threshold between use and development, but both feed into each other.
We didn't make the final selection in the IA Cluster call for tenders. Is this a hindrance to the development of AI in Montpellier?
No, because the IA cluster dynamic is still going strong. The players in the socio-economic world have reaffirmed their determination to pursue this dynamic. The Metropole is highly motivated, the Region voted for a very strong IA strategy this summer, and the financial backers are behind us... The next challenge will be to structure this dynamic and bring it to life. We're also fortunate to be working with the University Hospital, which is among the most dynamic and advanced in the field of AI and healthcare data processing. We have all the pieces of the puzzle in Montpellier, with a very, very favorable set of players, which I can't see anywhere else.