抖阴APP导航

Generative artificial intelligence (AI) 鈥� advanced machine learning systems like ChatGPT and Midjourney 鈥� is transforming how we learn, work and navigate daily life. Yet alongside AI鈥檚 growing influence come significant concerns: today鈥檚 models are energy-intensive and expensive, and can introduce biases into real-world decisions.

, PhD, an assistant professor and Schulich research chair in the Department of Electrical and Software Engineering at the , is leading a new to address these challenges. Ioannou will partner with , PhD, from the  (INRIA) and the  (ENS) in Paris to improve generative AI models.

鈥淚nternational partnerships strengthen the 抖阴APP导航 research community and accelerate innovation,鈥� says Dr. William Ghali, vice-president (research). 鈥淚mproving efficiency and fairness in AI is essential for the future of responsible and safe technology. We鈥檙e excited to see Dr. Ioannou leading this conversation on a global stage.鈥�

Tackling efficiency in modern AI

鈥淭he biggest challenge with current AI technology is the enormous amount of energy these models use,鈥� says Ioannou, who also leads the . 鈥淭hey鈥檙e highly inefficient to train and run.鈥�

Today鈥檚 generative AI systems demand vast computational and environmental resources. Training them can cost millions of dollars and require thousands of graphics processing units running for months, limiting access to who can use them. 

鈥淥nly major tech companies can afford to train these large models,鈥� says Ioannou.

While techniques exist to compress small AI models for faster operation, they do not work well with today鈥檚 large models. The research team will investigate why large models are difficult to compress. By understanding what happens inside these systems during training, the team aims to design new approaches that reduce energy use, cost and environmental impact without sacrificing performance.

Improving training methods and reducing costs can create broader access to AI research for scientists, startups and smaller companies, fostering a more collaborative and innovative landscape.

Understanding the impact on fairness

The team will also study how training and compression techniques influence fairness. Because AI models learn from real-world, human-generated data, they can reflect societal biases. While fairness in AI is an active area of research, few researchers have examined how efficiency efforts may unintentionally alter model behaviour. 

 鈥淎ll the methods we use to make models more efficient can also affect fairness. They can potentially make models less fair or change how they behave,鈥� says Ioannou.

Their project was funded by the (NSERC) and France鈥檚 (ANR), with supplemental support from , a Quebec-based non-profit consortium dedicated to responsible AI.

Opportunities for students and international collaboration

The project will create unique opportunities for 抖阴APP导航 and ENS students, including recruitment, international exchanges and joint training programs. 

At the end of the three-year grant, 抖阴APP导航 will host a workshop bringing together AI experts from Canada and internationally to share key findings and to foster new research collaborations.

Bridging theory and practice

A key strength of the project is the partnership with Simsekli, who specializes in the mathematical foundations of how AI systems learn from real-world data. His research complements Ioannou鈥檚 practical work designing efficient AI systems from an engineering perspective.

鈥淏ringing these strengths together is really exciting,鈥� says Ioannou. 鈥淭hat鈥檚 often where the biggest leaps in research come from.鈥�

鈥淎I already affects almost every part of people鈥檚 lives. Understanding how these models work, improving accessibility, and ensuring they treat people fairly is crucial for industry, research and society.鈥�