Adaptive Self-Supervised Learning for Generative AI: Enhancing Model Efficiency and Generalization in Low-Data Regimes

Abstract

Generative Artificial Intelligence (Generative AI) models have demonstrated remarkable capabilities in image synthesis, text generation, multimodal learning, and content generation. However, the training of such models generally requires large-scale labeled or weakly labeled datasets, substantial computational resources, and extensive training time. These requirements become particularly challenging in low-data regimes, where only a limited number of labeled or unlabeled samples are available. Conventional supervised learning methods may suffer from overfitting, poor representation learning, and limited generalization when training data are scarce.

This paper proposes an Adaptive Self-Supervised Learning (ASSL) framework for improving the efficiency and generalization capability of generative AI models under low-data conditions. The proposed framework combines adaptive pretext-task selection, curriculum-based difficulty scheduling, consistency regularization, representation learning, and adaptive sample selection. Instead of relying exclusively on manually labeled data, ASSL exploits the intrinsic structure of unlabeled data to learn robust representations. A dynamic task-selection mechanism selects suitable self-supervised tasks according to data characteristics and model learning status. Furthermore, a curriculum scheduler progressively increases task difficulty as the model improves.

The proposed framework is designed to reduce data requirements while maintaining competitive generative performance. The framework can be integrated with different generative architectures, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models. The evaluation framework considers generation quality, downstream classification accuracy, convergence speed, and computational efficiency. The proposed approach provides a scalable direction for data-efficient Generative AI in domains where large labeled datasets are unavailable.