Bringing AI to the Real World
Purchased from Gettyimages. Copyright.
Montréal is a recognized global leader in the Artificial Intelligence landscape. One of the city’s AI research hubs boasting a deep talent pool is, undoubtedly, l’École de technologie supérieure.
Among these talents is Mohammadhadi Shateri, Assistant Professor in the Systems Engineering Department. Upon joining ÉTS last summer, he took on a number of projects using AI and its various approaches—machine learning, deep learning, computer vision and reinforcement learning.
He is one of the lucky people who can declare: I coach Artificial Intelligence.
Ensuring Privacy and Security with AI
One of Shateri’s projects is related to data privacy and security in generative AI models. Since his PhD, he has been working on the subject with Professor Fabrice Labeau and Professor Pablo Piantanida, Director of the International Laboratory on Learning Systems.
Generative models have a plethora of practical applications in data privacy and the security of learning systems.
In domains where there is no access to real datasets due to privacy concerns, generative models compensate by generating real-looking data. Yet, they can leak information on the training dataset.
ÉTS professor Mohammadhadi Shateri
“We are changing the training process in the way that the models could continue to generate real-looking samples, and at the same time, minimizing the chances of leaking sensitive information,” explains Mohammadhadi.
As to data privacy, Shateri sees his mission as making sure that the AI models generate new representations of the data to keep their utility, but filter out sensitive information.
In his previous research, in collaboration with Hydro-Quebec, Shateri used a framework based on reinforcement learning as one solution for user data privacy in smart grids.
“When you consume electricity, your smart meter records the data and shares it with your energy provider. I trained a reinforcement learning agent which has access to physical resources at the user end (e.g., rechargeable batteries, electric vehicles, renewable energy sources), as well as in the grid, to decide how much energy it gets from each source at any given point in time,” clarifies Mohammadhadi. The agent hides the consumption pattern at the user end so that it appears to be random. Therefore, no one has access to the user’s private information.
Training AI for Real Life
His second project, in collaboration with Professor Éric Granger, Director of LIVIA (Laboratory of Imaging, Vision and Artificial Intelligence), is on Domain Adaptation in AI deep learning models.
Take, for example, AI models in autonomous driving. When a car is moving, a deep learning model should detect all sorts of objects using computer vision techniques. However, the model is trained mostly on computer-generated images of objects in a lab. In the real world, a shift is created in the domain of the data that the model is observing, which can degrade its performance.
To fill the gap between lab-trained data (source domain) and reality test data (target domain), models could be retrained, but this requires labelling the data, which takes time, energy and money.
“We took a different approach—to find a feature space shared across two domains, using the source domain knowledge available in the pretrained model,” says Shateri. The next step is to map the target domain into this feature space by minimizing a contrastive adaptation loss function.
Using AI in Biomedical Research
Another ongoing research with LIVIA focuses on AI’s biomedical application. The goal is to create an AI model for cardiotocography signal abnormality detection, used to analyze the fetal heart rate and ensure that the fetus receives enough oxygen in the womb.
“The model we developed, even unfinished, is surpassing all existing models. Nonetheless, we are working on more advanced algorithms to achieve a perfect analysis.”
The challenge is that researchers are looking for abnormal fetal heart rate signals when mostly good signals are available to them. The AI model, however, needs to learn the pattern of the problematic coding.
So, what do you do if you have access to highly imbalanced data?
One solution is to train an advanced generative model, for instance, by domain adaptation techniques applied to generative models to generate more abnormal heart rate data. “With enough data, our AI model will learn the pattern, and, eventually, improve its performance,” concludes Mohammadhadi.
“I am grateful to both ÉTS and LIVIA,” says Shateri. “They pave the way for you—you just have to show up and do the work. I am very lucky to be here, working on AI.”
