As artificial intelligence advances, the prevalent use of knowledge graphs (KGs) and retrieval-augmented generation (RAG) has become a much hyped cornerstone for structuring information and ensuring accuracy. However, these methods may also be holding AI back.
By aligning AI systems with traditional human ontologies — structured semantics frameworks grounded in human experience and logic — we may inadvertently limit AI's potential to discover new, purely correlation-based patterns that defy our rigid classifications.
The Constraints of Human Ontologies
Human ontologies are, by nature, restricted to what we already understand. Through scientific method, and human consensus seeking (which is less scientific than we might be comfortable with), we arrive at ontological truths.
KGs and RAG reinforce these structures by embedding human logic and categories into AI's operational framework. This grounding has clear benefits for accuracy and interpretability; it shows us new knowledge from within the box of what we accept as truth. However, it risks filtering out patterns or relationships that do not align with established categories. AI's inherent power lies in its ability to process vast amounts of data, beyond our biases or perspectives. Forcing it to operate within predefined human categories constrains this ability, keeping it from recognizing connections that are opaque or invisible to us.
Let me illustrate with a simple example. Existing knowledge graphs in food safety demonstrate the ontological path from flavour, through compounds (molecules) to human health impact. That makes sense. It's scientific. But by grounding Machine Learning to the ontology, we shut out pattern seeking between flavour and health. Now bear with me — I am not implying that there may be some new age non-scientific connection between flavour and health. That would not be scientific. But it is not absurd to speculate if flavour and the human experience of taste have complex interactions within the psychological and psychosomatic reality. Even though embodiment is a well-studied and scientifically accepted discourse, by limiting the AI to traditional human ontologies through KGs we are severely scoping pattern seeking within the available data sets.
Embracing Machine Ontologies Through Correlation-Based Discovery
Imagine AI models set free from rigid ontological frameworks and instead driven purely by mathematical correlation. Rather than forcing AI to follow existing human ontologies, we could allow it to generate machine ontologies — frameworks of knowledge that reflect patterns in the data itself. Such machine ontologies would not make sense immediately but could reveal connections we hadn't anticipated. Only ex-post facto, by mapping AI's conclusions back into our human frameworks, would these new ontologies gain meaning.
Why Machine Ontologies Could Be the Future of Knowledge Discovery:
- Unsupervised Correlation Mining: AI systems could identify non-obvious relationships that humans haven't yet conceptualised. By sifting through massive datasets without the constraint of human categories, they could flag patterns that are statistically significant but outside our conceptual reach.
- Hypothesis Generation: Once machine ontologies reveal novel correlations, they become a source of hypotheses that human researchers can then test through traditional scientific methods. This could accelerate discovery in fields where data is abundant but theory is lagging.
- Hybrid Ontologies for Human-AI Collaboration: The goal shouldn't be to replace human ontologies entirely but to create a dynamic interplay between human and machine frameworks. Imagine an AI system that operates with a machine ontology for discovery but translates its findings into human ontologies for communication and validation.
This article is based on Paul Devalier's reflections following EFSA's Data Readiness for AI Symposium.