Authors: Domas Buracas, Nathan Miller
[Presentation] [Conference-Style Report]
Abstract:
We seek to further validate the objective and theoretical principles introduced in Maximal Coding Rate Reduction (MCR2) by scaling it to the task of semantic segmentation in real world environments, where the number of labels per inference and size of datasets is significantly greater than the classification datasets STL10 and CIFAR100. Additionally, we show that applying a learned mask to filter high-frequency background pixels can greatly improve both the empirical and qualitative performance of MCR2. Finally, while the MCR2 objective does not appear to improve in learning domain agnostic features to assist with domain generalization, it improves upon standard supervised learning losses when the input space contains samples from many varied domains.
* This was completed as a final project for the graduate class EECS 208: Computational Principles for High-Dimensional Data Analysis