Networks trained on image datasets often exploit spurious correlations that are hard to detect, because the relevant biases are unknown in advance and the datasets are too large to inspect by hand. This project introduces an interactive web tool for concept-level auditing that combines CLIP’s semantic embedding space with SpLiCE, which decomposes image embeddings into sparse, human-readable concept representations.
Users compose queries as weighted combinations of natural-language concepts and retrieve matching images by nearest-neighbour search, with no model inference at query time. An enrichment panel then ranks concepts by how far they are overrepresented in the retrieved subset relative to the dataset baseline, and surfaced concepts can be fed back into the query builder.
The result is a hypothesis-free workflow: rather than specifying a suspected association up front, the user discovers candidate associations from what the retrieval returns. Demonstrated on MSCOCO, where it surfaces asymmetric gendered associations and spurious concept co-occurrences that would be difficult to anticipate manually.
Alex Fan, Arnav Gurudatt, Karsten Kropp, Gabriel López-Asiaín
University of Chicago