Comparison of semi-supervised learning algorithms

• المؤلف الرئيسي: Teo, Sheng Huai
• مؤلفون آخرون: Li Boyang
• التنسيق: Final Year Project
• اللغة: English
• منشور في: Nanyang Technological University 2023
• الموضوعات: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• الوصول للمادة أونلاين: https://hdl.handle.net/10356/172003
• المؤسسة: Nanyang Technological University

In this report, we conducted image classification experiments in a semi-supervised setting using three datasets of various sizes and content, CIFAR10, CIFAR100 and EuroSAT, with only 1000 samples labelled and the remaining as unlabelled for training. The performance and training duration of three SSL algorithms, MixMatch, FixMatch and FlexMatch, were compared. For CIFAR10 and EuroSAT, MixMatch achieved the highest accuracy of 95.8% and 93.4% respectively. Despite having the best performance for both datasets, MixMatch took most time to train, with an average of 27.5% longer than FixMatch, which has the shortest training duration. For CIFAR100, FixMatch obtained the best results for all four metrics obtained. FlexMatch was the next best, with an accuracy of 73.2%, and the other metrics having a difference of roughly 2.5% compared to FixMatch.