Towards Explainable Food Hazard Detection

As food systems grow more complex, ensuring food safety with transparent and interpretable AI is critical. This project introduces a novel neuro-symbolic approach for food hazard detection that prioritizes explainability without sacrificing performance, especially in low-data scenarios.

Key Contributions:

• Neuro-Symbolic Architecture: I co-developed a system that uses CoCo-Ex to extract key concepts from food recall notices and maps them to the ConceptNet knowledge graph. A Llama-3.1-8B-Instruct model was then used to filter and refine these concepts, creating a context-specific sub-graph.
• Explainable Classification: By leveraging the relationships and distances between concepts within the constructed knowledge graph, the system can classify hazards in a way that is inherently transparent and easy to interpret.
• Strong Low-Data Performance: While traditional neural-only baselines require vast amounts of data, this neuro-symbolic approach demonstrates superior performance in low-data settings, highlighting its potential for practical, real-world food safety applications where large datasets are not always available.

This work represents a significant step towards building more trustworthy and transparent AI systems for critical domains like food safety.

You can read the full paper here.