ICCV Daily 2021 - Friday

supervised loss plus an unsupervised loss. It is very hard not to overfit to the supervised loss, particularly when you have very few images. “ In supervised learning, you take data, and you create a test, then you have the computer take the test, you tell it the answers, and you have it keep taking the test again and again until it gets it right, ” he explains. “ I work a lot in education, and I would never try and teach a classroom of students this way! Yes, I’m sure I’ve had moments in my life where I’ve thought, wow, I’m actually learning something new while I’m taking this test, but it’s very rare that you learn by just taking a test over and over again. That was one of the inspirations for PAWS to remove this additional supervised loss and to integrate the label information directly into the algorithm . ” In terms of results, PAWS shows that integrating label information when it is available in a self-supervised learning framework can significantly improve performance and decrease training time . To the best of Mahmoud’s knowledge, PAWS is the first method to match the performance of fully supervised learning using only 10% of labels on ImageNet – so with 10 times less labels than fully supervised learning. With Mahmoud working both at Facebook and Mila , we have to ask him if he has had the pleasure of working with Yann LeCun or Yoshua Bengio . “ No, unfortunately I haven’t yet, ” he responds. What would he want to ask them if he did? “ For both actually, I would be really curious to get their opinion on the 9 DAILY ICCV Friday Mahmoud Assran