Generative Artificial Intelligence (AI) has emerged as a transformative technology, enabling the creation of realistic data and facilitating numerous applications in image generation, text synthesis, and predictive modeling. Despite significant progress, challenges persist in optimizing the performance and efficiency of generative algorithms, particularly in the context of Generative Adversarial Networks(GANs) and diffusion models.
This thesis addresses these challenges through a comprehensive investigation into evolutionary approaches for enhancing the effectiveness and efficiency of generative AI algorithms. Beginning with a thorough review of foundational concepts in generative AI, transfer learning, and knowledge distillation,the thesis identifies key research gaps and outlines the research questions driving the inquiry.
The primary research questions explored in this thesis are:
• How can the training time of EvolutionaryGANs(E-GANs) be shortened without compromising performance?
• How can knowledge distillation techniques be integrated with E-GANs to improve convergence speedand resource efficiency?
• What novel approaches can be developed to optimize the noises cheduler in diffusion models, thereby enhancing sampling speed and quality?
The study begins with an extensive review of foundational concepts in generative AI, transfer learning, and knowledge distillation, identifying critical research gaps and areas for improvement. Through a series of innovative methodologies and rigorous experiments, this thesis introduces several groundbreaking solutions. It presents the Partial Transfer Training-based E-GAN (PT-EGAN) and the Knowledge Distillation Evolutionary GAN (KDE-GAN), which significantly reduce training time and improve convergence speed of Evolutionary GANs while maintaining high-quality output. Additionally, a novel evolutionary approach is developed to optimize the noise scheduler in diffusion models,resulting in enhanced sampling efficiency and performance.
Overall, this thesis contributes to advancing the field of generative AI by introducing novel evolutionary approaches that improve both the performance and efficiency of state-of-the-art generative algorithms. These findings offer valuable insights for researchers and practitioners seeking to leverage generativeAI for various real-world applications, from image synthesis to predictive modeling.