UiT Machine Learning Group with paper accepted to CVPR 2021
We are proud to announce that the paper “Reconsidering Representation Alignment for Multi-view Clustering” by Daniel J. Trosten, Sigurd Løkse, Robert Jenssen and Michael Kampffmeyer, from the UiT Machine Learning Group / SFI Visual Intelligence was accepted in CVPR 2021.
CVPR http://cvpr2021.thecvf.com/ is the premier annual computer vision event and is known for its excellent quality and original research. In this competitive venue, only around 22% of submissions were accepted in 2021.
Among the accepted papers were “Reconsidering Representation Alignment for Multi-view Clustering” by Daniel J. Trosten, Sigurd Løkse, Robert Jenssen and Michael Kampffmeyer, from the UiT Machine Learning Group / SFI Visual Intelligence.
The paper presents novel methods for identifying groups in data gathered through multiple views or with multiple modalities, using deep neural networks. “We start the paper by identifying several drawbacks of current state of the art methods for deep multi-view clustering, related to the alignment of distributions of representations produced by the deep neural networks,” said Daniel J. Trosten, who is the lead author of the paper.
Armed with this new insight, the authors were able to design a simple but competitive baseline model for deep multi-view clustering, that avoids the alignment of distributions altogether. In addition, they identify a natural connection between multi-view learning and contrastive representation learning.
“By leveraging the connection between multi-view learning and contrastive learning, we were able to further improve on the baseline model by including components from recent self-supervised representation learning methods,” Trosten said.