Puyuan Liu wins CAIAC Best Masters Thesis Award

Puyuan Liu, a recent MSc graduate at the University of Alberta supervised by Amii Fellow and Canada CIFAR AI Chair Lili Mou, is being honoured with the 2023 Best Masters Thesis Award by the Canadian Artificial Intelligence Association (CAIAC).

"I was pretty surprised," says Liu, who graduated from the U of A in November. "I finished up, and then six months later, they say, 'You won this award; good job.' And then I was pretty grateful for the guidance and support of my supervisor."

The award is given out every year for "an outstanding masters-level thesis completed at a Canadian University in the field of Artificial Intelligence." Liu, who studies natural language processing (NLP), won for the paper "Non-Autoregressive Unsupervised Summarization with Length-Control Algorithms."

Liu says his research interests are broad, including robotics and reinforcement learning. But his current work in NLP comes from his interest in artificial general intelligence.

"If we want to build general AI or something, we need to somehow communicate with the AI, which will involve NLP for sure. So that's one of the things that interested me."

Liu's award-winning paper focuses on non-autoregressive text generation. Generative text programs use small groupings of letters and other characters called tokens. The models combine these tokens into the words and sentences that it outputs.

Traditional text generation models — such as Chat-GPT — use an autoregressive approach: it creates tokens one at a time, mashing them together to create a longer text. Non-autoregressive generators don't create tokens one at a time. Instead, they generate all the tokens simultaneously. Liu says this approach means the connections between the words might be slightly weaker, but the text is created much more quickly — sometimes 50 times faster than traditional methods.

"Basically, we're sacrificing a little bit of quality to get a large gain in efficiency," Liu says.

The paper proposed a non-autoregressive approach to summarizing larger pieces of writing. Liu and his co-authors found they could produce summaries with comparable accuracy much more quickly than traditional methods.

In addition to the award, Liu is invited to give a talk at CAIAC's conference from June 5 - 9.