Open set learning with augmented category by exploiting unlabelled data (open-LACU) [article]

Emile R. Engelbrecht, Johan A. du Preez
2022 arXiv   pre-print
Considering the nature of unlabelled data, it is common for partially labelled training datasets to contain samples that belong to novel categories. Although these so-called observed novel categories exist in the training data, they do not belong to any of the training labels. In contrast, open-sets define novel categories as those unobserved during during training, but present during testing. This research is the first to generalize between observed and unobserved novel categories within a new
more » ... learning policy called open-set learning with augmented category by exploiting unlabeled data or open-LACU. This study conducts a high-level review on novelty detection so to differentiate between research fields that concern observed novel categories, and the research fields that concern unobserved novel categories. Open-LACU is then introduced as a synthesis of the relevant fields to maintain the advantages of each within a single learning policy. Currently, we are finalising the first open-LACU network which will be combined with this pre-print to be sent for publication.
arXiv:2002.01368v4 fatcat:pbyuj7lbcjc3jjldp7oih3slye