Toward Deep Semi-Supervised Continual Learning: A Unified Survey for Scalable and Adaptive AI

The integration of Deep Semi-Supervised Learning (DSSL) with Continual Learning (CL) holds significant promise for advancing artificial intelligence systems capable of learning from limited labeled data while continuously adapting to new tasks. This review explores recent progress in combining DSSL and CL, referred to as Deep Semi-Supervised Continual Learning (DSCL), focusing on the potential to develop models capable of learning efficiently in dynamic environments through both labeled and unlabeled data while mitigating catastrophic forgetting. Our analysis highlights several critical applications including image classification, cybersecurity, and natural language processing, while also identifying key challenges that prevent its broader adoption in real-world scenarios. Integration challenges such as catastrophic forgetting, handling noisy, unlabeled, and imbalanced data, and managing the stability-plasticity trade-off are discussed in detail. Moreover, the importance of open-world learning, lightweight architectures for on-device learning, enhanced scalability, and interpretable models is emphasized to ensure DSCL’s applicability in real-world, high-stakes domains. However, the current reliance on benchmark datasets, while valuable for evaluation, may limit generalization to complex tasks like medical imaging. To address these challenges, future research should prioritize leveraging domain-specific datasets to enhance real-world applicability, integrating transfer learning for better adaptability, and developing domain-agnostic frameworks and task-free continual learning. Additionally, exploring techniques like Reinforcement Learning from Human Feedback (RLHF) could enhance interpretability and trustworthiness. By addressing these gaps, DSCL can evolve to provide more flexible, scalable, and reliable solutions, contributing significantly to the development of adaptable and intelligent systems across diverse domains.