�鶹��Ƶ

Six researchers at the University of Toronto with expertise in areas such as machine learning, medicine, engineering and statistics are joining the Vector Institute.

They’re joining a highly accomplished team of world-class researchers who helped launch the Vector Institute in March, including �鶹��Ƶ's  Emeritus Geoffrey Hinton; Associate Professor and leader of Uber’s Advanced Technologies Group Raquel Urtasun; and Deep Genomics founder and electrical and computing engineering Professor Brendan Frey.

“My top priority since Vector’s launch has and continues to be to build a collaborative, talented team,” said one of the original Vector Institute co-founders, Richard Zemel, a �鶹��Ƶ computer science professor and the research director for Vector. “I am thrilled that these individuals will join a diverse and growing team of faculty committed to making Vector a global leader in AI.”

The new members are:

• David J. Fleet, �鶹��Ƶ professor of computer science, department of mathematical sciences at the University of Toronto Scarborough
• Anna Goldenberg, �鶹��Ƶ assistant professor of computer science (computational biology group) and scientist in the Genetics and Genome Biology Lab at SickKids Research Institute
• Frank Rudzicz, �鶹��Ƶ assistant professor of computer science (status) and scientist at the Toronto Rehabilitation Institute-UHN
• Jimmy Ba, assistant professor of �鶹��Ƶ computer science, joining �鶹��Ƶ in fall of 2018
• Murat Erdogdu, joint appointment as �鶹��Ƶ assistant professor of  computer science and statistical sciences, joining �鶹��Ƶ in fall of 2018
• Marzyeh Ghassemi, joint appointment as �鶹��Ƶ assistant professor of computer science and medicine, joining �鶹��Ƶ in fall of 2018

Originally from Turkey, Erdogdu (left) is coming to �鶹��Ƶ from Microsoft Research New England, where he is a postdoctoral researcher. He has a PhD in statistics and a master’s degree in computer science from Stanford University, as well as bachelor degrees in electrical engineering and mathematics from Boğaziҫi University.

His specialty is the design of optimization algorithms for machine learning models, such as deep learning models and recommender systems. By using more efficient algorithms, model training times can be significantly reduced from weeks (depending on the size of the datasets) to just hours, enabling researchers to efficiently test and select the best model for the problem at hand.

“I’m really looking forward to designing efficient algorithms for real-world problems,” Erdogdu said, adding he’s an avid hiker and camper and is looking forward to exploring the city and surrounding countryside when he’s not in the lab.

“When I visited Canada, I found it extremely friendly and it was really diverse,” he said.

Invited to speak on campus last March, he said he was blown away by the “energetic” and collaborative nature of the computer science department and all of the activity on the downtown Toronto campus. He was also invited to attend the official launch of the Vector Institute.

“The University of Toronto is one of the best universities in my area of research, which is artificial intelligence, machine learning and optimization,” he continued. “It will be a pleasure to be around experts like Geoff Hinton and Radford Neal – I think it’s safe to say they changed the world. There are also lots of young faculty that I’d be very excited to work with and collaborate with.

“And that’s why I chose the University of Toronto and Vector.”

Ba (left), whose research focuses on the development of learning algorithms for deep neural networks, is returning to �鶹��Ƶ – this time in computer science – after completing his undergraduate, master’s and PhD here in the Edward S. Rogers Sr. Department of Electrical & Computer Engineering.

He was supervised at �鶹��Ƶ by some leading researchers in machine learning, including Hinton, Frey and Ruslan Salakhutdinov, who is now a computer science professor at Carnegie Mellon University.

Among his many accomplishments, Ba developed the Adam Optimizer, one of the go-to algorithms to train deep learning models. He was also one of the first students from a Canadian institution to win a Facebook PhD Fellowship, and in 2015 his team achieved the highest place among academic labs in the image caption generation competition at the Conference on Computer Vision and Pattern Recognition. He is currently a postdoctoral researcher at MIT.

Ghassemi (left), a recent PhD graduate from MIT,  is a visiting researcher with Alphabet’s Verily and a part-time postdoc at MIT. Her hiring at �鶹��Ƶ signals a new partnership between computer science and medicine: She is the first joint hire in computational medicine.

“The University of Toronto computer science department has established experts in machine learning,” said Ghassemi, “as well as really exciting junior faculty doing work in important areas of machine learning like Raquel [Urtasun] and driverless vehicles."

She has focused her research on machine learning applications in health care. This means using machine learning as a tool in her work exploring clinical data with algorithms to “predict interesting and important human risks” and anticipate patient needs and decrease mortality rates, she explained.

For example, clinical data can be mined to determine what kind of patients in the intensive care unit will require a ventilator or a blood cell transfusion. It can also predict a patient's length of stay and their risk of death within a year of leaving care.

Ghassemi is also interested in using machine learning in non-invasive patient monitoring. She’s worked on detecting voice disorders that impact a subject's ability to speak using an electromechanical device called an accelerometer, coupled with a machine learning algorithm.

Instead of placing a camera in a patient’s throat that detects polyps and other damage to strained vocal chords, the accelerometer is simply taped to their neck.

“We were able to non-invasively detect which patients had nodules, which was very exciting,” she said. Her clinical collaborators at the Massachusetts General Hospital Voice Center are currently exploring the use of a cellphone prompting system based on findings of the non-invasive accelerometer. When subjects are in danger of straining their chords, they are sent an alert.

In addition to its machine learning expertise, the ability to work with “world-class clinical collaborators” at �鶹��Ƶ’s nine partner hospitals was another important draw for Ghassemi.

“All of the people I met at the Faculty of Medicine were excited to work on clinically meaningful problems,” she said. “They really wanted to understand how we could apply machine learning techniques and develop new algorithms to be useful in a clinical setting. That combination of a fantastic technical school and a collaborative, world-class clinical environment is pretty rare.”