CVPR Daily - Friday

That would be great! I am Italian. We will perform a duet in Arabic and Italian! Yes! I would love that. Back to your work now. Can you tell us more about it? Self-supervised learning is learning from data without labels. Traditionally, in machine learning, you have some input and some labels, and then you train a model that maps the input to the labels. We humans do not do the mapping. Annotating images is very expensive. In some domains, like the medical domain, experts are needed. There are millions of unlabeled data out there. That is why there was the urge to move to self-supervised learning. But until recently, self- supervised learning was not outperforming supervised learning. In April 2021, we proposed Group Masked Model Learning (GMML) using a transformer. You take an image, corrupt some parts, and ask the network to find what was corrupted. For example, you have an image of a bird, and you corrupt most of it, but the face of the bird is still visible. If the network can reconstruct it, it means it understands the notion of the image. That was the main idea. For the first time, we outperformed supervised learning and convolutional neural networks. The community picked it up, people were talking about it, it was shared on Twitter, and my friends were mentioning me in posts. It was great. Congratulations! That is so nice. Yes, I thought it was just the start of something. Then we sent it to a journal, and we got a rejection. It ignored the idea we were proposing and how powerful it was and only cared about one thing: we did not use big data sets. We are a small group with limited resources; we are not a tech giant. Most researchers are small groups with limited resources. But we said that is fine, and we will try to do something to move forward. Later, we were disappointed to find that other researchers and tech companies had started using our idea without acknowledging us as the first to propose it. It was heartbreaking. We proposed this in April 2021 and put it on arXiv. Everyone can see it. What can younger scholars learn from your experience? At some point, I thought that I should just quit if I am not acknowledged. I worked hard in the office day and night to provide 24 DAILY CVPR Women in Computer Vision Friday