Browsing by Author "Engelbrecht, Emile"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Open-set learning with augmented category by exploiting unlabelled data (Open-LACU)
    (Stellenbosch : Stellenbosch University, 2024-03) Engelbrecht, Emile; Du Preez, Johan ; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: Neural network classifiers provide scalable means to analyse categorical patterns within datasets. However, current machine learning policies fail to consider certain nuances developed in real-world applications. The vast number of patterns represented in certain datasets and the continual collection of new data means classifiers must be aware of the observed-novel category and the unobserved novel category. To address these challenges, this dissertation combines semi-supervised learning and novelty detection into a single learning framework called open-set learning with augmented category by exploiting unlabelled data or Open-LACU. Although Open-LACU requires further development, we show and argue that Open-LACU classifiers will have reduced annotation cost, improved practicality and enhanced safety. Semi-supervised learning trains models using partially labelled datasets to reduce annotation costs. Novelty detection ensures classifiers are able to separate all data samples outside the domain of interest for enhanced safety. When working with partially labelled datasets in a domain where novel patterns exist, several inconsistencies appear in existing literature. More specifically, there is no distinction between those novel patterns which are unrepresented during training but appear during testing, and those novel patterns that are represented in unlabelled training data. Considering the unique properties of these different novel category types, we argue that classifiers must generalise these separately. In Open-LACU, classifiers must generalise 1) those K > 1 number of source categories for which labels are provided, 2) an additional K + 1’th observed-novel category for those novel patterns in the unlabelled training data, and 3) an additional K + 2’nd unobserved-novel category that encapsulates all those novel patterns unobserved during training but seen during testing. To introduce Open-LACU, we pursue several objectives that integrate different learning frameworks. For each of these integrating steps, we experiment on small-scale vision datasets to simulate different categorical scenarios. Our results both confirm the feasibility of Open-LACU and reveal several insights into the challenges that future research must address.