Multi-Depth Concept Extraction for Post-Hoc Vision Encoder Explanation

NAVE enables non-invasive model inspection: The method works on already-trained vision models without requiring fine-tuning or architectural modifications, making it practical for auditing existing AI systems deployed in real-world applications.

Multiple levels of detail reveal how models process information: By extracting concepts from multiple depths within the encoder, NAVE can show explanations at different levels of granularity—from simple features to complex patterns—helping stakeholders understand what the model "sees" at different stages of processing.

Works across different model types: The method is architecture-agnostic, functioning with both traditional CNNs and newer Vision Transformers, which means a single explanation approach can be applied to diverse vision models rather than requiring custom solutions for each type.

Extracted concepts align with human understanding: The paper demonstrates that NAVE's identified concepts correspond to meaningful semantic features in the input, suggesting the explanations reflect genuine patterns the model uses rather than artifacts, which is critical for trustworthy AI auditing.

Enables comparative analysis of model design choices: By using NAVE as an inspection tool, researchers can analyze how different training strategies and architectures affect what models learn internally, supporting more informed decisions about which models to deploy responsibly.