������Ƶ

As issues around the usage and ethics of artificial intelligence (AI) continue to evolve, new ������Ƶ graduate is ready to address those challenges. 

“The public should embrace AI as a catalyst for positive change,” says Zhang, who graduates this month with his PhD from the in the Faculty of Arts & Science.

“Ensuring that AI systems align with human values and remain under human control becomes increasingly critical. Addressing AI safety is one of the most important and impactful problems we face today.”

Zhang intends to be part of the solution. He applied to U of T for his PhD several years ago, inspired by the “godfather of AI” ������Ƶ,  emeritus in the department. 

As a PhD student, Zhang taught many courses at U of T. His study of theoretical foundations and practical algorithms for machine learning have already been adopted by major AI players, including Google Brain, DeepMind and OpenAI.

Ahead of convocation, Faculty of Arts & Science writer David Goldberg spoke with Zhang about his research and how U of T prepared him for a future career in AI.

Why was U of T the best place to earn your PhD?

I was captivated by the immense potential of deep learning, so U of T was an obvious choice for me given its leadership in this area. ������Ƶ and his students shocked the world with their results on ImageNet in 2012 with AlexNets, a neural network architecture which started a golden age for deep learning. Furthermore, the prospect of collaborating with the rendered U of T even more special.

How do you explain your work with AI to people outside your field?

I focused on developing neural network models and algorithms that excel in fast training, robust generalization and accurate uncertainty estimation. With neural networks often comprising millions or billions of parameters, the challenge lies in understanding effective optimization techniques for these networks. I am also exploring the phenomenon of why neural networks have such impressive generalization abilities. And finally, another key aspect of my research was investigating whether neural networks can possess awareness of their knowledge gaps.

How is your work with AI going to improve life for the average person?

My research on neural network training dynamics holds significant importance in the realm of large language models and AI research. These models used in programs such as ChatGPT, which have become ubiquitous in our daily lives, play a vital role in various applications. For example, they assist us with translation, enhance our essay writing and serve as virtual assistants to address our queries.

There’s controversy surrounding some of the ways AI is being used – why do you think people need to embrace its potential?

The public should embrace AI as a catalyst for positive change because it enhances efficiency and productivity across industries, automates mundane tasks allowing for more meaningful work and improves problem-solving and decision-making through data analysis. In addition, it augments human capabilities and drives innovation while also helping us to address societal challenges like inequality and sustainability.

Embracing AI responsibly ensures transparency, accountability and ethical considerations, unlocking AI's potential for positive impact in our society. I think the public should also be involved in regulating AI, as AI systems could be very powerful and misuse of them could lead to catastrophic consequences. 

What career path will you pursue after graduation – and how will your U of T education help you excel?

I will work as an AI researcher in industry, focusing on large language models. My education at U of T has equipped me with extensive knowledge in deep learning and artificial intelligence. Under the guidance of my advisor, Associate Professor , and collaboration with colleagues, I have gained valuable insights into neural network training dynamics and AI safety.

This expertise enables me to enhance the efficiency of large language model training while ensuring alignment with human values. My PhD work on understanding neural network training with a noisy quadratic model has already been used by many big industrial labs (including Anthropic, DeepMind, OpenAI) in training the latest models.

What advice do you have for people considering their PhD – in your field or beyond?

I recommend everyone maintain a curious mindset and pursue their passions. Curiosity is vital for scientific progress. It is also crucial to remain open-minded and committed to lifelong learning. Our field is rapidly evolving, rendering knowledge from just a few years ago potentially outdated, so continuous learning is essential.