Start Date: November 1, 2024
Vasileios Sioros
Generative AI models have shown impressive abilities in areas like image generation and natural language processing. However, these models often depend too much on learned patterns, which can lead to outputs that seem convincing but lack factual accuracy. Issues such as hallucinations in language models and artifacts in generated images reveal important challenges in creating high-quality, reliable outputs. This research aims to overcome existing limitations by developing knowledge-guided and rule-validated generative AI systems that produce reliable outputs based on both data-driven techniques and rule-based knowledge. This hybrid approach merges the statistical strength of neural networks with the structured logic of symbolic AI. Neuro-symbolic AI can be classified by how deeply neural and symbolic components are integrated. Earlier models primarily use neural networks for pattern recognition and symbolic AI for logical reasoning, while more advanced models embed symbolic reasoning within neural architectures. Additionally, this research will explore alternative mechanisms like graph neural networks (GNNs) and attention mechanisms, which may improve relational reasoning and contextual understanding in generative models. The primary focus of this study aligns closely with the natural sciences. For instance, research has introduced novel models that generate stable materials by learning from the data distribution of existing stable structures while incorporating energy minimization and bonding preferences as guiding principles. Additionally, these models include validation mechanisms to ensure that the generated crystals adhere to physical laws. Key areas of focus include: Conducting a literature review of existing knowledge-guided generative AI models. Developing methods to guide machine learning models in mitigating issues like hallucinations in language and artifacts in images. Implementing automatic validation mechanisms for generated content with minimal human intervention. Establishing end-to-end pipelines that integrate both knowledge guidance and validation to enhance reasoning within generative models.