Mixup and Metric Learning in Out-of-Distribution Detection

Hutor, Oleksandra

Home
→
Students Research & Project Works | Роботи студентів
→
Факультет прикладних наук
→
Бакалаврська програма "Комп'ютерні науки"
→
2023
→
View Item

dc.contributor.author	Hutor, Oleksandra
dc.date.accessioned	2024-02-14T11:52:25Z
dc.date.available	2024-02-14T11:52:25Z
dc.date.issued	2023
dc.identifier.citation	Hutor, Oleksandra. Mixup and Metric Learning in Out-of-Distribution Detection / Oleksandra Hutor; Supervisor: Giorgos Tolias; Ukrainian Catholic University, Department of Computer Sciences. – Lviv: 2023. – 52 p.: ill.	uk
dc.identifier.uri	https://er.ucu.edu.ua/handle/1/4435
dc.language.iso	en	uk
dc.title	Mixup and Metric Learning in Out-of-Distribution Detection	uk
dc.type	Preprint	uk
dc.status	Публікується вперше	uk
dc.description.abstracten	In this work, we analyse the problem of out-of-distribution detection, which includes inlier classification and outlier detection, and the use of two methods for it, i.e., Mixup and metric learning, as well as their combination. Inspired by the use of Mixup between inliers and outliers in MixOE (Zhang et al., 2023), our first objective is to identify the key ingredients of their method and to combine it with seven other Mixup techniques. We also change the size and diversity of the auxiliary outlier dataset, which is used for training. We find that in the fine-grained OoD setting, where outliers come from the same domain as inliers, mixing only the label with the one that comes from the uniform probability distribution without mixing the inlier and outlier training samples has a similar performance as mixing both pairs of training samples and labels. At the same time, in the coarse-grained settings, where outliers come from a completely different domain than inliers, the more outliers are used for mixing, the better the detection performance is. Our second objective is the investigation of metric learning in the form of triplet loss for the same setup with different types of triplet combinations, some of which are created using Mixup. In the coarse-grained settings, the triplet combination with only inlier classes performs better than such combinations with outliers, mixed or not. Finally, we combine MixOE and our metric learning approach to show that, for some datasets, the detection performance in the coarse-grained settings is comparably larger than the previous best result, MixOE.	uk