IMPROVING SELF-SUPERVISED REPRESENTATION LEARNING IN MOCO V2 WITH QUEUE OPTIMIZATION

• Main Author: Jofandi, Gugun
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87791
• Institution: Institut Teknologi Bandung

This research is motivated by the need to improve the performance of selfsupervised learning models, particularly the Momentum Contrastive version 2 (MoCo v2) architecture. This study aims to develop a more robust and accurate MoCo v2 model by adding a K-Nearest Neighbors (KNN) mechanism to the queue. The research method employed is experimentation, modifying the MoCo v2 architecture through KNN integration to filter the queue and select stronger positive representations. The results demonstrate that MoCo v2 + KNN achieves a 5% accuracy improvement to 87.7% on the CIFAR-10 dataset compared to the MoCo v2 baseline model. Utilizing KNN in filtering the queue proves effective in selecting more discriminative positive representations, thereby enhancing model performance. Furthermore, MoCo v2 + KNN exhibits better resilience to large queue sizes, overcoming MoCo v2's sensitivity to excessive queue size. This research also utilizes strong data augmentation, which has been shown to effectively increase model robustness in previous studies. In conclusion, adding KNN to the MoCo v2 for queue filtering successfully improved accuracy by 5% compared to the MoCo v2 baseline, resilience to queue size, and the overall performance of the self-supervised learning model. The research highlights the potential of incorporating KNN into MoCo v2 for advancing self-supervised learning in Object Detection.