Search
  • Artificial Intelligence
  • Profile
  • Artificial Intelligence

Sanja Fidler is taking data to the next level with AI-assisted annotation tool

by Krista Davidson Sep 9 / 20

Sanja Fidler is helping researchers scale their AI innovations by creating an AI-assisted tool that can rapidly and accurately label large amounts of data.

A Canada CIFAR AI Chair and faculty member of the Vector Institute, Fidler specializes in computer vision, specifically in 2D and 3D object detection. She is also an associate professor at the University of Toronto and a director of NVIDIA’s Toronto AI Lab. 

Traditionally, data labelling, or annotation, is a laborious process undertaken by humans. Fidler says it takes a long time and can become incredibly tedious and prone to fatigue-related errors. However, labelling the data is essential for training robust algorithms.

“Data is really the lifeblood of machine learning,” she explains. “Without data, there is no machine learning. In a lot of machine learning domains, especially in computer vision, labelling data becomes a huge bottleneck,” she says, adding that in some cases data collection and labelling make up the majority of a project’s time and expense. According to Fidler, some companies may opt for less data and therefore poorer performing algorithms in order to meet project/prototype deadlines.

Fidler and her team of students at the University of Toronto were inspired to create an annotation platform that could help industry and academia scale their products and services. 

“We decided to create a software that was going to take all of these techniques that we’ve developed out of the lab and incorporate them into an AI-assisted labelling platform that researchers and companies can actually use,” she says.

The application automatically completes the tasks of segmenting pixels and labeling objects, which are important for many different applications, including medical imaging self-driving cars. 

Fidler’s research could significantly speed up this important process. “If it took you a few weeks to label a million images, we envision you could be able to now label ten to a hundred million images at the same time in the future,” she says.

Fidler was awarded the Connaught Innovation Award for her work in AI-assisted annotation in 2020.