Open-world learning under dataset shift

• Main Author: Srey, Ponhvoan
• Other Authors: Philipp Harms
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Science::Mathematics
• Online Access: https://hdl.handle.net/10356/172053
• Institution: Nanyang Technological University

Conventional classification models in machine learning are imposed with strict constraints, limiting their implementation in real-world scenarios. Datasets encountered in the wild naturally contain instances of both known and unknown classes. Furthermore, at test time, the data is frequently drawn from a different distribution or domain compared to the training data. In this project, we introduce an end-to-end framework that simultaneously handles the open-world nature and the shift in domain between training and test data. At the core, we adapt domain adaptation techniques to the open-world setting, and propose to minimise the uncertainty of predicting the unlabelled data, thereby improving model generalisation. We demonstrate the effectiveness of our method on challenging benchmark datasets, with an improvement of 106%, 9%, and 25% in overall accuracy on ImageNet, VisDA and CIFAR-100, respectively. We further test our method under different settings, highlighting its robustness.