CAFD: Concept-Aware DNN Fault Detection using VLMs
Researchers have developed a new method called Concept-Aware Fault Detection (CAFD) to identify errors in Deep Neural Networks (DNNs). CAFD integrates various data sources, including a novel "Concept Failure Ratio" derived from Vision-Language Models (VLMs). This ratio uses VLMs to extract textual concepts from images and assess their link to DNN failures, providing valuable semantic information. CAFD demonstrated superior performance over existing methods, achieving an average 18.3% improvement in Fault Detection Rate across multiple DNN models and datasets. AI
IMPACT Enhances reliability of AI systems by improving methods for detecting faults in DNNs.