Computer Vision News - November 2022

4 Best Paper Award Xingjian Zhen is a PhD student at the University of Wisconsin-Madison, advised by Vikas Singh. Zihang Meng is a Research Scientist at Meta AI in New York. He recently graduated from UW-Madison with the same advisor. Rudrasis (Rudra) Chakraborty is a Senior Research Scientist at Butlr, a start-up in California. He graduated with a PhD in Computer Science in 2018. Together, they just scooped the Best Paper Award at ECCV 2022 for their work, introducing the traditional statistical domain of distance correlation into deep learning. They are here to tell us all about it. ON THE VERSATILE USES OF PARTIAL DISTANCE CORRELATION IN DEEP LEARNING With so many neural networks available, choosing which one to use for a given task can be difficult. There may be several different options,whichposes an important question: How do you know if Network A is better than Network B ? The community often addresses such a simple question in an unnecessarily complex way. In approaching this work, Xingjian, Zihang, and Rudra believed that no one had looked for a simple answer. “ We’re conditioning one network on another network, ” Xingjian tells us. “ Most of the time, when people are comparing networks, they care more about the performance or accuracy, but questions regarding the information remaining in the network have not been well studied. We’ve borrowed the partial distance correlation method from the statistical domain to remove the information from one network off another pre-trained network. We compare the remaining information to see if it’s still meaningful regarding our data. Then we can say that one network contains more information than another. ” Rudra adds: “ Correlation is very simple. Everyone can understand the relationship between two things. That’s why we borrowed something from the statistics textbook, which is well understood and well known, to answer this question. ” Once the team had figured out the mathematical side, incorporating that into the engineering part was more challenging due to the large gap between theory and practice. B E S T PAPER ECCV