Zheda Mai (MASc Student), Jihwan Jeong (PhD Candidate), Prof. Scott Sanner and team member Hyunwoo Kim (LG Sciencepark) won first prize in all categories in the CLVision Challenge held at the 2020 virtual Computer Vision and Pattern Recognition (CVPR) conference. Their entry beat out 79 other teams including some top Continual Learning/Computer Vision teams in the world from both industry and academia.
Continual learning is the ability of a system to continuously learn and adapt from the data it collects over time This requires the system to be able to learn the current task without forgetting the ability to perform well in all previously seen tasks. It poses a huge challenge due to device storage limits that don’t allow for the option to repeatedly review all previously seen data. Continual learning is often applied in low power and memory devices such as cell phones and will play an increasingly integral role in Artificial Intelligence systems that continually learn and adapt to the data they take in.
The CLVision Challenge asked competitors to submit continual learning solutions for three different challenge tracks: New Instances (NI), Multi-Task New Classes (Multi-Task-NC) and New Instances and Classes (NIC). Teams had the option to submit to one or all three of these tracks and the competition was held in two separate phases, pre-selection and final evaluation. The LG & University of Toronto team easily made it into the final evaluation phase and ranked number one in all three categories to win first prize overall.
“I was very excited to take part in this competition as it allowed us to apply novel continual learning solutions and provided a comprehensive evaluation on a shared hardware platform for a fair comparison,” Zheda said, “Each solution was evaluated across a number of metrics, including accuracy, ram usage, running time, etc. This made the competition more challenging since we need to strike a balance between all the metrics.”
Zheda and his team presented a continual learning method called batch-level experience replay with review in their entry. In this method a memory buffer stores samples of each task that is completed. When the system is asked to preform a new task, it reviews and learns from the samples stored in the memory buffer. They used this method and its variants on all three tracks set forth in the CLVision Challenge and achieved impressive results.
To learn more about the CLVision Challenge and the University of Toronto & LG Science Park submission you can review their paper online.
-Published July 8, 2020 by Lynsey Mellon, lynsey@mie.utoronto.ca
