Vol. 2, Issue 2, Part A (2025)

Self-supervised learning (self-supervised learning (SSL)) has emerged as a transformative paradigm in artificial intelligence, enabling models to learn meaningful representations from vast quantities of unlabelled data without manual annotation. This research investigates and compares key self-supervised learning (SSL) paradigms contrastive, predictive, and hybrid frameworks to determine their relative efficacy across varied domains, including natural image classification and medical imaging. Using benchmark data sets such as CIFAR-10, ImageNet-1K, and CheXpert, multiple self-supervised learning (SSL) architectures SimCLR, MoCo v2, BYOL, SimSiam, Barlow Twins, VICReg, and JEPA were evaluated through standardized experimental protocols. Statistical analyzes, including ANOVA and Tukey’s post-hoc tests, were employed to validate performance differences. Results reveal that hybrid and predictive-regularized models like VICReg and JEPA consistently outperform contrastive-only approaches, achieving superior Top-1 accuracy and AUC-ROC across data sets while maintaining greater representational stability and generalization. The integration of predictive and contrastive objectives with variance-covariance regularization proved especially effective in minimizing representation collapse and enhancing feature diversity. The study concludes that self-supervision, when designed through hybrid learning objectives, offers a scalable and domain-agnostic approach to representation learning, reducing dependency on annotated data sets. Practical recommendations include adopting multi-objective loss functions, designing domain-specific augmentation strategies, and employing adaptive optimization schedules to ensure stable learning and cross-domain transferability. Overall, this research reinforces the paradigm shift from supervised dependency toward autonomous, data-efficient, and generalizable learning systems capable of advancing real-world artificial intelligence applications in healthcare, remote sensing, and industry.