This text is part of the special Research section
Understanding why bees die, predicting police stress, detecting professional burnout among nurses … the National Scientific Research Institute (INRS) has infrastructures enabling cutting-edge research to be carried out with theartificial intelligence, at the service of local businesses.
“We provide the knowledge that companies need”, sums up Tiago Falk, full professor at INRS and scientific manager of the Multimedia / multimodal signal analysis and improvement laboratory (Musae Lab).
“Usually when we talk about artificial intelligence, you have a model and you are looking for data to apply. You modify the model to apply the information it finds, ”says Falk. But the INRS proceeds in the opposite way. “We have the data, we know all the details, all the nuances. And then we build the models, ”he continues about information collected in realistic environments, not simulated or artificially obtained data.
This therefore makes it possible to create artificial intelligence models that can be applied in practice, based on information collected in the field.
The importance of context
“When you know where the data is coming from, you are able to create models that are right, that work,” says Falk. The context is also crucial to analyze a situation with precision.
The researcher cites as an example a project carried out with the National Police School of Quebec (ENPQ), which was interested in stress among students and officers. Using smartwatches and cell phones, certain body reactions were measured to determine for signs of stress.
Some artificial intelligence models detect stress as soon as the heart rate increases. “But these models were built using data that was manipulated in the lab,” says Falk. Is it because the police officer is running or because he is stressed? Is he tired? You need to know how to separate the change in heart rate caused by running and the change caused by stress. “
The importance of context is also valid in research on the signs of burnout among hospital nurses. “If a nurse is stressed in the cafeteria, it has much less consequences than if she is in an intensive care unit,” he says.
The same applies to bees, which are the subject of studies by the INRS to determine the causes of their death. “Most go outside to collect pollen. If we see the bees coming out of the hive telling themselves that it is an emergency and that they are simply foraging, we must know, ”he illustrates.
Fight against bias
But data entered into an artificial intelligence model can also lead to biased results. “A machine cannot learn models just on the basis of the data it has had,” says Falk. For example, an algorithm could draw conclusions with an inherent bias on an individual’s gender or ethnicity. INRS is currently working with sociology researchers who support them in collecting information in order to eliminate this type of discrimination.
Thus, precaution is in order, particularly in a project aimed at detecting signs of stress and mental health problems in people living in remote regions. “You have to be careful, you can have many biases. Depression itself has many faces, depending on your background and your history, ”warns Mr. Falk.
Reduce the GHG impact
Energy-intensive, artificial intelligence? An article published in 2019 by researchers at the University of Massachusetts concludes that it does. Thus, the process of training large AI models can emit up to five times the greenhouse gas (GHG) emissions equivalent of an average American car over its entire lifespan. “These are models that have to be trained for many weeks, running on hundreds of computers,” says Falk.
INRS is therefore working to find a way to reduce these emissions, in particular by using optical technology and light. “With a fraction of the power, we can achieve the same performance as with processors,” says the researcher. This will therefore allow artificial intelligence to be more sustainable by having fewer energy needs, which will generate fewer GHG emissions, he hopes.
Watch video
Reference-feedproxy.google.com