ECE Graduate and Postdoctoral Studies Student Perspective Stories

The landscape of deep-learning models has undergone significant changes in the last few years, characterized by a steady increase in the size of these models. Leading the way in this trend are large language models, which have become increasingly prevalent and influential in the field of artificial intelligence. The emergence of such large models has brought with it a host of system-level challenges, particularly in relation to their training and inference. Training these models across distributed computational resources remains an open problem, while the deployment of large models for inference poses its own set of challenges. My research interest centres at the intersection of architecture and systems, with a particular focus on addressing the challenges posed by large deep-learning models.

The environmental and social costs of producing and disposing of electronics is a serious global problem, with concerning impacts on human health in developing countries, on the environment and on carbon emissions. Electronic products, and consequently e-waste, contain various toxic elements and pollutants, which pose a major threat to both ecosystems and to human health when discarded. Manufacturing these devices is also highly carbon intensive, with the embedded carbon of electronics often accounting for the majority of their total carbon footprint. In computer engineering, problems are typically only viewed through a technical lens, and only addressed using technical solutions, but this is a socio-technical problem, involving technical design, economic models of production and consumption, consumer behaviors and policy. To address growing environmental and ethical problems in computer engineering, we need to take a holistic approach, developing a richer understanding of the broader social context to inform technical solutions and bringing technical expertise into community collaborations, education and policy work. My research investigates how consumer behaviors and attitudes can inform the design of more sustainable, repairable electronics, how sustainability literacy (specifically around the environmental impacts of computing) and practical computer skills can be advanced through community engagement and education and how to develop sustainable alternatives to existing models of personal electronic design and consumption. A key part of my research involves community-based education. Firstly, through workshops and short-term courses, I’d like to help the broader community develop the computer literacy and technical skills required to engage in electronic repair and to better understand and articulate related concerns and needs. Secondly, I want to develop strategies for bringing community perspectives and meaningful sustainability education into the university-level electrical and computer engineering curriculum. Thirdly, through partnership and consultation with community members and groups, my work investigates the challenges and concerns of diverse groups when using personal electronics and engaging in maintenance and repair activities, in order to develop technical guidelines and policy recommendations.

My research involves determining different analytical techniques to solve for and approximate end-to-end distribution times of stochastic directed acyclic graphs. This has many applications in high performance computing and cloud computing. The specific process we are attempting to improve is determining the end-to-end time distribution of a task, where the task can be broken into subtasks and distribution of the subtasks is known. The challenging aspect of the problem is when the subtasks are dependent on each other, with some requiring others to finish before they can start. In this problem the time is often very difficult to calculate exactly, which leads to the need for good approximations.

Machine learning (ML) has achieved remarkable performance in many tasks like image classification, and already seen great prospects in many real-world applications. ML can facilitate precision medicine to empower clinical decision making, maneuver the driving vehicle without human intervention. On the other hand, existing ML technology is also brittle and prone to failure that could entail critical consequence: (1) a hardware mistake can translate to a software failure that causes an ML model to exhibit unexpected behavior, such as misclassifying a stop sign as a speed limit sign; (2) the ML model can also be easily fooled by the inputs that have been tampered with; and (3) struggle to provide both high quality of service and strong privacy protection when the model is trained on sensitive data. This engenders serious concern on the trustworthiness of ML. My research concentrates on three prominent challenges (reliability, security and privacy) in trustworthy ML and advocates a multi-faceted solution to improve the reliability, security and privacy of ML to fully deliver the promise of benefits of ML.

My research is about protecting mobile phone users from malware. In particular, I aim to protect users from malware that uses logic bombs, a technique by which malware hides its malicious intents from malware analysis. This requires developing efficient code analysis techniques for monitoring app behaviours, malware detection technique that can operate on a slice of the code which contains the untested behaviour, and develop hardware accelerators that can run the analysis efficiently.