Zohreh Aghababaeyan, a computer science PhD student at the Faculty of Engineering, is conducting research to ensure the reliability of AI systems and autonomous technology.
Zohreh’s research is redefining the standards for trustworthy AI and making systems more secure and dependable. Her work not only enhances current AI technologies but also lays the foundation to seamlessly integrate AI into everyday life.
We spoke with Zohreh to learn more about her groundbreaking research and future goals.
What motivated you to pursue research in your field?
I’m drawn to deep neural networks (DNNs) and AI because of their incredible potential to change our world. AI is everywhere now, but the big question we face is how to make it reliable and trustworthy, especially when it’s used in important areas of our lives.
Every day, I’m motivated to work on making AI systems more efficient. By solving the challenge of reliability and safety, we can really start to use AI’s full power in all areas of life, making things easier and better for everyone.
Can you tell us more about your research?
In the domain of autonomous technology, where data generation by a single Tesla vehicle can amount to 11 GB per day, the challenges associated with labelling and effectively selecting essential inputs for deep neural networks (DNNs) evaluation are significant. Relying on traditional metrics such as accuracy to assess DNN models is insufficient, as these metrics merely quantify the number of mispredictions without delving into their underlying causes or the similarities among them.
To address these challenges, my PhD research develops a test selection strategy that focuses on diversity and uncertainty to manage large-scale data and complex labelling in DNN evaluation. Using an automated framework, it identifies key inputs that reveal significant DNN faults without needing internal access. The DNN fault detection framework I introduced also moves beyond accuracy, offering deeper performance insights. This approach, validated through extensive experiments, proves superior in boosting DNN safety, reliability and comprehension, redefining AI evaluation standards.
My research targets the crucial requirement for accurate testing and evaluation of AI models underscored in safety-critical sectors such as medicine and autonomous driving. Through the development of pioneering DNN testing methodologies, my work not only elevates the reliability of AI systems but also reduces the resource cost of data labelling. These advancements foster greater confidence in AI technologies, paving the way for the rollout of safer and more reliable applications across various industries.
What are your career aspirations?
As a PhD student and researcher who works on trustworthy AI, my goal is to make AI more reliable and accessible for everyone, ensuring its use is met with confidence and efficiency. I aspire to contribute to cutting-edge AI research that breaks new ground in reliability, driving innovation that helps integrate AI smoothly and securely into everyday life.
Sustainable and resilient infrastructure
Zohreh Aghababaeyan’s research project secured her first place in the sustainable and resilient infrastructure category at the 2024 Engineering and Computer Science Graduate Poster Competition held at the Faculty of Engineering. Her project, titled Elevating DNN Reliability: A Comparative Study of Test Selection Metrics, is supervised by Professor Lionel Briand, who holds a Canada Research Chair in Intelligent Software Dependability and Compliance.
There are five areas of focus in the engineering faculty’s research, including sustainable and resilient infrastructure research.
This article is part of our series on the winners of the 2024 Engineering and Computer Science Graduate Poster Competition.
